Fuzzing with AFL workshop
Materials of the "Fuzzing with AFL" workshop by Michael Macnair (@michael_macnair).
This workshop introduces fuzzing and how to make the most of using American Fuzzy Lop, a popular and powerful fuzzer, through a series of challenges where you rediscover real vulnerabilities in popular open source projects.
The first public version of this workshop was presented at SteelCon 2017 and it was revised for each of BSides London 2019, BSides Bristol 2019, and GrayHat 2020 (most notable change in this revision was a switch to afl++).
GrayHat published a recording of a remote version of the workshop on YouTube - this was created for a real-time workshop audience, but you can follow along at your own pace as long as you don't mind skipping a few pauses and ignoring references to Discord.
The presentation suggests when to attempt the different challenges in this repository, and the video provides a
- 3-4 hours (more to complete all the challenges)
- Linux machine
- Basic C and command line experience - ability to modify and compile C programs.
- Docker, or the dependencies described in
- quickstart - Do this first! A tiny sample program to get started with fuzzing, including instructions on how to setup your machine.
- harness - the basics of creating a test harness. Do this if you have any doubts about the "plumbing" between afl-fuzz and the target code.
- challenges - a set of known-vulnerable programs with fuzzing hints
- docker - Instructions and Dockerfile for preparing a suitable environment, and hosting it on GCP if you wish. A prebuilt image can be pulled from ghcr.io/mykter/fuzz-training.
See the other READMEs for more information.
Challenges, roughly in recommended order, with any specific aspects they cover:
- libxml2 - an ideal target, using ASAN and persistent mode.
- heartbleed - infamous bug, using ASAN.
- sendmail/1301 - parallel fuzzing
- ntpq - fuzzing a network client; coverage analysis and increasing coverage
- date - fuzzing environment variable input
- cyber-grand-challenge - an easy vuln and an example of a hard to find vuln using afl
- sendmail/1305 - persistent mode difficulties
The challenges have HINTS.md and ANSWERS.md files - these contain useful information about fuzzing different targets even if you're not going to attempt the challenge.
Most of the challenges also have an ANSWERS-libFuzzer.md file, for if you want to try out using LLVM's libFuzzer. These are brief descriptions of the differences for libFuzzer, and should be read alongside the afl docs (.md files).
All of the challenges use real vulnerabilities from open source projects (the CVEs are identified in the descriptions), with the exception of the Cyber Grand Challenge extract, which is a synthetic vulnerability.
The chosen bugs are all fairly well isolated, and (except where noted) are very amenable to fuzzing. This means that you should be able to discover the bugs with a relatively small amount of compute time - these won't take core-days, most of them will take core-minutes. That said, fuzz testing is by definition a random process, so there's no guarantee how long it will take to find a particular bug, just a probability distribution.
- The afl docs/ directory
- Ben Nagy’s “Finding Bugs in OS X using AFL” video
- The afl-users mailing list
- The smart fuzzer revolution (talk on the future of fuzzing): video / slides
- A categorized collection of recent fuzzing papers (there are a lot!)
- The Fuzzing Book - broad coverage of fuzzing
- More challenges from an EkoParty workshop
- Introduction to triaging crashes
- Google's ClusterFuzz and Microsoft's OneFuzz