Follow @Openwall on Twitter for new release announcements and other news
[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Date: Fri, 21 Nov 2014 08:27:14 +0100
From: Gynvael Coldwind <>
Cc: Mateusz Jurczyk <>
Subject: Re: Fuzzing project brainstorming

> > If it came from a mutation-based fuzzer, the original (not-mutated) sample
> > can be useful too.
> You mean the closest non-crashing parent (in case there is a chain of samples as in AFL)?
Generally yes, but I guess in cases where the bug was caused by
multiple changes (vs a fully properly formatted base sample) in
different areas and different generations both the base sample and
previous-generation sample would be useful to have (as well as the
crashing sample ofc).

> And while we are at it, would you mind describing your experience in case of ffmpeg.
> Your blogpost -- -- gives only high level review of the
> work. The fuzzer and specific methods of fuzzing seems to be proprietary. That's fine.
> But perhaps you can describe other sides of the work:
(+j00ru as he might have additional insight or patches to the text I'm
about to write)

Well, the fuzzing/mutation methods were pretty much described in the blogpost:
"simple algorithms (such as bitflipping, swapping bytes, truncating
the files and so forth)"
Later in the process we got to some more interesting stuff, though not
super magical either.

> - how did you deduplicate crashes (full stacktrace, some frames only or some other way);
We used de-ASLRed call stack traces as keys. It wasn't perfect, but we
did a lot of iterations so even if we missed something because of that
we would still find it in the next iteration (hehe OK, I realize
that's not totally true; it's like saying "oh, we forgot to claim the
prize from this lottery ticket; no matter, we'll just win the lottery
again next week", but it was "good enough" I would say).

> - how did you decide which issues are security-sensitive and which are not;
We didn't, mostly due to the sheer amount of the unique (deduplicated)
crashes we were getting. In ffmpeg case (as well as other cases that
were/are getting tons of unique crashes) we kinda delegate this to the
teams that are maintaining the project - I know it's far from perfect
from the maintainers point of view, but we just didn't have the
manpower to analyze everything, especially that ffmpeg isn't the only
thing we're fuzzing.
So, generally in this case the ffmpeg team has done all the work in
this area (though we might occasionally say 'this looks bad' and 'this
look meh').

> - how did you requested CVEs (for which issues, which info was required);
No magic here either - the issue selection for CVEs was done by ffmpeg
maintainers as well (I believe in some minor cases we might also point
out crashes that looked bad and would probably need a CVE). So
basically after the fixes were done (and ffmpeg team has always fixed
stuff super fast - we were always really amazed by their dedication to
this, just wow! I wish some commercial vendors would be as fast) the
ones that looked exploitable (as in "exploitable into RCE", not just
"triggerable" / "DoS") from the "fix perspective" got a CVE.
As for info required, it's best to ask the MITRE guys, but afair the
information we sent them consisted of a link to the commit/patch that
also had some very brief text describing the bug.

> - (if you know) how security fixes were released by ffmpeg.
You would have to ask the maintainers for details about it, but afair
there is no special release process (please correct me if I'm wrong).
The fixes are committed to the repository and later get a CVE. This
doesn't sound perfect, but please note that in our blog post we did
mention that the recommendation for ffmpeg is a privilege-separated
(sandbox) environment anyway. TBH my personal opinion is that this
applies to any media player with C/C++ parsing code out there - I'm
yet to spot one that has no bugs.

> This kind of questions.
> Given the sheer number of findings you probably did everything automatically?
On our side most of the stuff (fuzzing, deduplication, reproduction,
initially also crash report generation) was done automatically (we
used 2,000 cores or so for this).
The manual part was coding the fuzzing system / mutators / etc of course ;)

Gynvael Coldwind

Powered by blists - more mailing lists

Please check out the Open Source Software Security Wiki, which is counterpart to this mailing list.

Confused about mailing lists and their use? Read about mailing lists on Wikipedia and check out these guidelines on proper formatting of your messages.