Concurrency strikes MSIE (potentially exploitable msxml3 flaws)
A while ago, apparently angry with Larry Seltzer, I penned a quick
write-up on the possible issues with race conditions triggered by
asynchronous browser events (such as JavaScript timers) colliding with
synchronous content rendering:
http://seclists.org/vulnwatch/2006/q3/0023.html
This is in principle similar to signal handling bugs. I gave an example of
a seemingly exploitable flaw in Firefox (see MFSA2006-59 report for more
details), but did mention that other browsers are unlikely to be immune.
Today, inspired by Brian Krebs' report on MSIE's stellar track of security
response that we all owe to responsible disclosure, I thought it would be
a brilliant idea to test MSIE for the same class of problems (they had
half a year to take notice of my original rant).
Hey, and - no peeking! - guess what happened?
A quick demonstration of how MSIE succumbs to such problems would be to
prepare an XML file that contains a bunch of nested tags (10-1000 is
fine), then display it in IFRAME, repeatedly disrupting the rendering
process with a Javascript timer that forces the frame to be reloaded every
50-100 miliseconds or so.
After just a couple reloads, MSIE will freeze, then crash in a random
manner in the vicinity of msxml3 module. I observed seemingly harmless
NULL pointer dereferences, writes to bogus addresses, reads from
unallocated memory, and various other signs of memory corruption typical
of such race conditions. The exact mode of crash depends on precise timing
and the contents of browser memory (previously / concurrently displayed
pages, contents of the rest of the document), but this is obviously well
within the control of a determined attacker.
As such, it is my guess that although (as with all race conditions) this
would be fairly hard to exploit remotely in a reliable way, it is within
the realm of possibility.
A quick "vanilla" but reasonably reliable demo that will probably freeze
then crash your browser on a NULL pointer dereference (or sometimes a
mangled target pointer on REP MOVSW or something along these lines, if you
came there from some other website) can be found at:
http://lcamtuf.coredump.cx/iediex/iediex.html
...try using the 'genie.sh' utility provided in the same directory to
generate more elaborate test cases that, combined with additional contents
of the pages will likely trigger more interesting memory corruption
scenarios.
/mz