It’s been almost five years since I created SharpFuzz, the only .NET coverage-guided fuzzer. I already have a blog post on how it works, what it can do for you, and what bugs it found, so check it out if this is the first time you hear about SharpFuzz:
A lot of interesting things have happened since then. SharpFuzz now works with libFuzzer, Windows, and .NET Framework. And it can finally fuzz the .NET Core base-class library! The whole fuzzing process has been dramatically simplified, too.
Not many people are aware of all these developments, so I decided to write this anniversary blog post and showcase everything SharpFuzz is currently capable of.
The list of bugs found by SharpFuzz has been growing steadily and it now contains more than 80 entries. I’m pretty confident that some of the bugs in the .NET Core standard library would have been impossible to discover using any other testing method:
- BigInteger.TryParse out-of-bounds access
- Double.Parse throws AccessViolationException on .NET Core 3.0
- G17 format specifier doesn’t always round-trip double values
As you can see, SharpFuzz is capable of finding not only crashes, but also correctness bugs—the more creative you are in writing your fuzzing functions, the higher your chances are for finding an interesting bug.
SharpFuzz can also find serious security vulnerabilities. I now have two CVEs in my trophy collection:
- CVE-2019-0980: .NET Framework and .NET Core Denial of Service Vulnerability
- CVE-2019-0981: .NET Framework and .NET Core Denial of Service Vulnerability
If you were ever wondering if fuzzing managed languages makes sense, I think you’ve got your answer right here.
Initial SharpFuzz usage instructions were unnecessarily complicated and full of error-prone, manual steps. I decided to completely rewrite these guidelines and also write a PowerShell script that correctly configures all parameters behind the scenes, making the whole fuzzing process way simpler. You no longer have to choose which assembly to instrument, which memory limit to use, or how to configure the execution timeout. You only need to:
- Create the fuzzing project and write your fuzzing function.
- Create one or more test cases.
- Run the fuzz.ps1 script like this:
pwsh scripts/fuzz.ps1 YourFuzzingProject.csproj -i Testcases
In addition to AFL, SharpFuzz now supports libFuzzer as a fuzzing engine. If you are interested in technical implementation details, check out the libfuzzer-dotnet repository. If you just want to start using libFuzzer, usage instructions are available here. AFL and libFuzzer have similar capabilities, so on its own, using libFuzzer instead of AFL doesn’t really bring you any benefits (it might be slightly easier to use, though, because you only have to download a single binary and you are ready to go). However, libFuzzer support in SharpFuzz has unlocked some exciting new possibilities: native Windows support and fuzzing .NET framework libraries.
After I finished writing the libFuzzer driver, I had completely forgotten about it. I joined Microsoft later that year, and a few months into my Microsoft career, I stumbled upon an internal fuzzing channel, where I discovered that there was a team in Microsoft that was working on porting my SharpFuzz libFuzzer driver to Windows! They not only wrote the port, but they were also happy to open source it. Huge thanks to Joe Ranweiler and the MORSE team for what they have done—I’m deeply thankful for their contribution, and I can’t emphasize enough how important it was for the SharpFuzz users. Native Windows support has made fuzzing much more accessible: despite .NET Core’s cross-platform success, most .NET users are probably still on Windows.
Everyone is moving to .NET Core these days. But the migration can be slow, and there are many libraries that might never be ported to .NET Core. With libFuzzer and Windows support, SharpFuzz can now fuzz .NET Framework libraries, too! I didn’t have to do anything special to enable this feature, though: SharpFuzz has been targeting .NET Standard since the beginning. The only thing needed to support .NET Framework was the fuzzing engine that could run on Windows, and as you know from the previous section, libFuzzer satisfies that requirement.
Fuzzing .NET Core standard library
My initial attempts to fuzz the .NET Core standard library (aka base-class library or BCL) failed, because I thought that modifying mixed-mode assemblies was impossible (and all official .NET assemblies are built as mixed-mode assemblies). But I really wanted to find bugs in the .NET standard library, so I ultimately figured out how to fuzz it by doing some godawful hacks like downloading the IL-only, nightly packages from a special NuGet feed used by the .NET team, and building the .NET runtime repo in managed-only mode (my eyes are now bleeding from trying to read the usage instructions I wrote about this procedure back in the day). Unfortunately, this solution was short-lived, because the .NET team stopped publishing the IL-only packages. They also changed the way .NET runtime was built, and I couldn’t figure out how to build it in managed-only mode again. This marked the end of my .NET Core BCL fuzzing attempts.
…until a few weeks ago. I was sitting in a coffee shop with a renewed enthusiasm for building the .NET runtime when it suddenly dawned on me: why do I even think that instrumenting mixed-mode assemblies is impossible? What if I unlearned that belief and started from scratch? A few minutes later, I realized that the impossible was not only possible (at least on Windows), but it took me only a few moments to implement. I still don’t really know how mixed-mode assemblies work, but I do know that I can easily strip the native code from them, and nothing will break. It’s funny how sometimes you need to wait for years for things to fall into place.
If you want to get started with fuzzing the .NET BCL, you can find plenty of examples in my dotnet-fuzzers repo (contributions are welcome). Everything you need to get started is there: fuzzing projects, dictionaries, and PowerShell commands. For example, if you want to fuzz the Uri.TryCreate method, you simply need to clone the repo and run the following script:
.\fuzz.ps1 ` -project .\src\UriFuzzer\UriFuzzer.csproj ` -corpus .\src\UriFuzzer\Testcases\ ` -targetDlls System.Private.Uri.dll
Unfortunately for us bug finders, the .NET team has been doing a really great job in recent years, so discovering new bugs in .NET Core has become more difficult. Every time I thought “there must be a bug here”, I was disappointed because the code turned out to be remarkably well designed. Well done, .NET team, I hope you are happy for basically ruining my life.
It makes me really happy to see so many awesome people participating in the SharpFuzz development. I’m profoundly grateful to Joe Ranweiler and the MORSE team, Günther Foidl, Ulrich Fourier, George Pollard, and p4fg for their contributions. Some people have written blog posts about SharpFuzz, and it was even a topic of a research project (it’s in German, though). Thank you all for being such a great community!
After taking a break from fuzzing for several years, I’m again actively working on SharpFuzz. There is still plenty of important and fun work in .NET fuzzing (for example, ASP.NET Core fuzzing and structure-aware fuzzing), so expect more posts in the future. In the meantime, enjoy fuzzing!
Big thanks to my loyal sidekick Milica Miljkov for her continuous support in my blogging efforts.