Perhaps not surprisingly, there appears to be a vulnerability in how
Microsoft Internet Explorer handles (or fails to handle) certain
combinations of nested OBJECT tags. This was tested with MSIE
6.0.2900.2180.xpsp.040806-1825 and mshtml.dll 6.00.2900.2873
xpsp_sp2_gdr.060322-1613.

At first sight, this vulnerability may offer a remote compromise vector,
although not necessarily a reliable one. The error is convoluted and
difficult to debug in absence of sources; as such, I cannot offer a
definitive attack scenario, nor rule out that my initial diagnosis will be
proved wrong [*]. As such, panic, but only slightly.

Probably the easiest way to trigger the problem is as follows:

  perl -e '{print "<STYLE></STYLE>\n<OBJECT>\nBork\n"x32}' >test.html

...this will (usually) cause a NULL pointer + fixed offset (eax+0x28)
dereference in mshtml.dll, the pointer being read from allocated but still
zeroed memory region.

The aforementioned condition is not exploitable, but padding the page with
preceeding OBJECT tag (and other tags), increasing the number of nested
OBJECTs, and most importantly, adding bogus 'type=' parameters of various
length to the final sequence of OBJECTs, will cause that dereference to
become non-NULL on many installations; then, a range of other interesting
faults should ensue, including dereferences of variable bogus addresses
close to stack, or crashes later on, when the page is reloaded or closed.

[ In absence of sources, I do not understand the precise underlying
  mechanics of the bug, and I am not inclined to spend hours with a
  debugger to find out. I'm simply judging by the symptoms, but these
  seem to be indicative of an exploitable flaw. ]

Several examples of pages that cause distinct faults in my setup (your
mileage may and probably WILL vary; on three test machines, this worked as
described; on one, all examples behaved in non-exploitable 0x28 way):

  http://lcamtuf.coredump.cx/iedie2-1.html (eax=0x0, instant dereference)
  http://lcamtuf.coredump.cx/iedie2-2.html (bogus esi on reload/leave)
  http://lcamtuf.coredump.cx/iedie2-3.html (page fault on browser close)
  http://lcamtuf.coredump.cx/iedie2-4.html (bogus esi on reload/leave)

Well, that's it. Feel free to research this further. This vulnerability,
as requested by customers, is released in strict observance of the Patch
Wednesday & Bug Saturday policy.

[*] The ability of the attacker to document the attack scenario probably
    doesn't matter for those who pretend to care; cryptic "hi" to
    Secunia and their standards of conduct.