Compromising pictures of Microsoft Internet Explorer!

[Michal Zalewski <lcamtuf@xxxxxxxxxxxx>]

Synopsis:
---------

  Well, not really. Instead, at the risk of boring you to death, I'd like
  to report on a casual 30-minute experiment I've conducted of recent.
  This experiment resulted in identifying a potential remote code
  execution path in Microsoft Internet Explorer, plus some other bugs, and
  should be a good starting point for further testing of other browsers or
  similar programs.

Discussion:
-----------

  You might remember the 'mangleme' affair, where various browsers were
  subjected by yours truly to a trivially constructed malformed HTML
  crash-course - all that in order to find exploitable input handling flaws.
  Back then, MSIE performed admirably compared to other browsers (although
  did not escape some embarassment when ned@felinemenace found the
  infamous IFRAME bug that way):

    http://lcamtuf.coredump.cx/mangleme/gallery/

  Of recent, I decided to try something completely different and radically
  new, without having to do any actual work. I used the same META REFRESH
  auto-test framework to check for image decompression and parsing flaws
  (JPEG, GIF, PNG), as opposed to making fun of HTML renderers.

  I used a simple index.cgi script (attached, though hardly noteworthy) to
  dynamically generate a page that references ten just as dynamically
  created images. These images were prepared by running a test set of
  pictures (some regular ones, and several pathological cases created with
  ImageMagick) through a slightly modified version of my old afx utility.

  Surprisingly, it is MSIE and its proprietary JPEG decoder (apparently
  not shared with other Windows components?) that performed embarassingly
  poor this time. Results below.

Vulnerability examples:
-----------------------

  NOTE #1: As with mangleme, this list of problems is most certainly NOT
  exhaustive, and performing longer tests or improving the technique
  would most likely result in additional findings.

  Several MSIE crash sample files from that 30-minute run are available
  at:

    http://lcamtuf.coredump.cx/crash/

  Note that these may produce different results depending on program
  versions, plugins and configuration. Tested with WinXP Pro PL
  2600.xpsp2.050301-1526 SP1, MSIE PL 6.0.2800.1106, up-to-date.

  mov_fencepost.jpg - on most platforms, causes a crash due to mov
    destination fencepost error after going past allocated memory, or
    after accessing a bogus address such as 0x27272727. The destination
    address appears to be controllable (i.e. changing the file or
    displaying other data before or along with this image alters it).
    My bets are that this is exploitable for remote execution.

  cmp_fencepost.jpg - here, causes a crash due to a very similar cmp
    fencepost (no write). Not necessarily exploitable for remote code
    execution, unless code execution path can be affected later on.

  oom_dos.jpg - usually causes a OOM crash. Less interesting, unless
    you like to punish people who borrow your pictures for their blogs.

  random.jpg - causes mov fencepost of CPU consumption + crash. Didn't
    investigate in much detail.

  NOTE #2: MSIE comes with no sources, and reverse engineering is naughty.
  I didn't examine the renderer to see what went wrong; I see unbounded,
  user-dependent memory accesses, and that spells trouble.

Vendor notification:
--------------------

  It is my experience that reporting and discussing security problems with
  Microsoft is a needlessly lengthy process that puts too much burden and
  effort on the researcher's end, especially if you just have a crash
  case, not a working exploit; hence, they did not get an advance notice.

Bonus (OT)
----------

  Since piggyback request smuggling and fooling proxies and filters is a
  popular new pastime, some of you might find it entertaining to have a
  look at how various applications differ in handling duplicate instances
  of HTTP/SMTP message/NNTP headers that are, in common perception,
  "supposed to" occur only once.

-- 
------------------------- bash$ :(){ :|:&};: --
 Michal Zalewski * [http://lcamtuf.coredump.cx]
    Did you know that clones never use mirrors?
--------------------------- 2005-07-14 00:29 --

      http://lcamtuf.coredump.cx/silence/

#!/bin/bash

echo "Content-Type: text/html"
echo

ID="timg-$$-$RANDOM-$RANDOM"

rm -f timg-* AFX.log

cat <<_EOF_
<HTML>
<HEAD>
<META HTTP-EQUIV="Refresh" content="0;URL=/">
</HEAD>
<BODY>
_EOF_

CNT=0

for i in img/*; do
  CNT="$[CNT+1]"
  FNAM="$ID-$CNT"
  EXT=`echo $i | cut -d. -f2`
  ./afx-loc -p 1 -i 100 -m RANDOM -s 60000 <$i 2>$FNAM.$EXT >>AFX.log
  echo "Test $CNT - <IMG SRC=\"$FNAM.$EXT\"><BR>"
done