Notes on fuzzing ImageMagick and GraphicsMagick

ImageMagick and GraphicsMagick are two popular libraries for manipulating images. GraphicsMagick is a fork of ImageMagick that diverged well over a decade ago. OSS-Fuzz provides continuous fuzzing for high impact open source projects. In December, 2017 Paul Kehrer and I worked to add ImageMagick to Google’s OSS-Fuzz, and in February, 2018 we added GraphicsMagick.

Both ImageMagick and GraphicsMagick had been widely fuzzed and audited before this. Hanno Böck 1 observed: “In the past it was pretty easy to fuzz bugs in imagemagick, but after some review by Google most of them have been fixed and these days there are at least no more trivial to find fuzzing issues.” Despite this, within hours of adding each project to OSS-Fuzz it was finding security issues. Between the two projects it has found more than 425 security issues of various severities 1, and it continues to occasionally find new ones.

Given the gaping chasm between what was expected and the massive success of OSS-Fuzz on ImageMagick and GraphicsMagick I thought it would be helpful to review what factors I thought were contributing to OSS-Fuzz finding so many vulnerabilities and other bugs:

Scale OSS-Fuzz leverages Google’s massive server farms to bring serious compute to bear on fuzzing. The last time I attempted to measure, it looked like they were running something like 30 CPU cores per fuzz target. Further, while lots of past fuzzing of ImageMagick and GraphicsMagick was done using AFL, OSS-Fuzz uses libFuzzer which gives the potential for higher executions per second. This gives them the ability to find bugs that take many iterations to show up. Finally, for ImageMagick and GraphicsMagick we generate one fuzz target per decoder. Each has more than 100 decoders (including many that leverage third party libraries such as libpng or libjpeg), and OSS-Fuzz runs them all, this is an amount of compute that is well beyond what’s accessible to most folks.

Continuous OSS-Fuzz builds an updated copy of the project every day, and it runs indefinitely. Compared to other fuzzing which might spin up a big EC2 instance and then run for a week, this is a huge advantage. It means that new bugs are caught as they are introduced and as time passes the amount of CPU dedicated to fuzzing each target will continue to climb. It also means if there’s something blocking progress (e.g. an OOM) as it gets fixed, the fuzzer will incorporate the fix and proceed.

Automated OSS-Fuzz automatically files tickets for each bug it finds, and tracks when they are fixed. This means crashers never slip through the cracks and reproducers never get lost; the issue tracker always knows the state of every bug it has encountered. It can also easily catch if a bug regresses.

MSAN Most fuzzing these days happens with ASAN, which means things like use-after-free and buffer-overflows are caught. However, ASAN doesn’t catch use of uninitialized memory. As Chris Evans demonstrated, uninitialized memory in ImageMagick can be used to exfiltrate secret data in memory. MSAN catches use of uninitialized memory, but unfortunately, using it is kind of a pain: every library you use, including libc, needs to be compiled with MSAN. OSS-Fuzz automatically handles building things with MSAN.

Improvement In addition to security issues, OSS-Fuzz also files bugs on memory leaks, timeouts, and out-of-memory issues. When folks are doing their own fuzzing, they often won’t bother to file bugs from these categories, because their objective is to find security issues, and manually filing bugs can be a pain. However, bugs in these categories can make fuzzing far less efficient, reducing the chance that the fuzzer will find security issues. By encouraging projects to fix these types of issues, OSS-Fuzz leads to projects being more efficient to fuzz, which helps find more vulnerabilities.

Conclusion

If you’re fuzzing (or considering fuzzing) an open source library, work with the maintainers to include it in OSS-Fuzz. Google will even pay you a bounty for the integration. It’s pretty clear to me that OSS-Fuzz will deliver better results than fuzzing on your own, making us all more secure.

I’d like to extend a huge thank you to the ImageMagick and GraphicsMagick teams, who were supportive of our efforts to integrate their projects into OSS-Fuzz, and who took on the lion’s share of the work: resolving both the vulnerabilities that were reported, as well as the other bugs.

And finally, I’d be remiss if I didn’t point out that basically every vulnerability class that OSS-Fuzz finds is a product of memory unsafe languages, like C and C++. While fuzzing makes these projects more secure, it’s not a substitute for using languages that don’t cause thousands of vulnerabilities. When we’re finding hundreds and thousands of vulnerabilities that all have a preventable root cause, it’s time to reconsider what we’re doing.


  1. I don’t want to pick on Hanno, this post is about how good OSS-Fuzz is, not how bad everyone else is. He just happened to have a quote that captured this well. ↩︎