AFL++ is a superior fork to Google’s AFL - more speed, more and better mutations, more and better instrumentation, custom module support, etc.

American Fuzzy Lop (AFL) was developed by Michał “lcamtuf” Zalewski starting in 2013/2014, and when he left Google end of 2017 he stopped developing it.

At the end of 2019, the Google fuzzing team took over maintenance of AFL, however, it is only accepting PRs from the community and is not developing enhancements anymore.

In the second quarter of 2019, 1 1/2 years later, when no further development of AFL had happened and it became clear there would none be coming, AFL++ was born, where initially community patches were collected and applied for bug fixes and enhancements. Then from various AFL spin-offs - mostly academic research - features were integrated. This already resulted in a much advanced AFL.

Until the end of 2019, the AFL++ team had grown to four active developers which then implemented their own research and features, making it now by far the most flexible and feature rich guided fuzzer available as open source. And in independent fuzzing benchmarks it is one of the best fuzzers available, e.g., Fuzzbench Report.

The definition of the terms whitebox, graybox, and blackbox fuzzing varies from one source to another. For example, “graybox fuzzing” could mean binary-only or source code fuzzing, or something completely different. Therefore, we try to avoid them.

The Fuzzing Book describes the original AFL to be a graybox fuzzer. In that sense, AFL++ is also a graybox fuzzer.

We compiled a list of tutorials and exercises, see tutorials.md.

A program contains functions, functions contain the compiled machine code. The compiled machine code in a function can be in a single or many basic blocks. A basic block is the largest possible number of subsequent machine code instructions that has exactly one entry point (which can be be entered by multiple other basic blocks) and runs linearly without branching or jumping to other addresses (except at the end).

function() {
  A:
    some
    code
  B:
    if (x) goto C; else goto D;
  C:
    some code
    goto E
  D:
    some code
    goto B
  E:
    return
}

Every code block between two jump locations is a basic block.

An edge is then the unique relationship between two directly connected basic blocks (from the code example above):

              Block A
                |
                v
              Block B  <------+
            /        \       |
            v          v      |
        Block C    Block D --+
            \
              v
              Block E

Every line between two blocks is an edge. Note that a few basic block loop to itself, this too would be an edge.

To stop afl-fuzz, minimize it’s corpus and restart you would usually do:

Control-C  # to terminate afl-fuzz
$ afl-cmin -T nproc -i out/default/queue -o minimized_queue -- ./target
$ AFL_FAST_CAL=1 AFL_CMPLOG_ONLY_NEW=1 afl-fuzz -i minimized_queue -o out2 [other options] -- ./target

If this improves fuzzing or not is debated and no consensus has been reached or in-depth analysis been performed.

On the pro side: * The queue/corpus is reduced (up to 20%) by removing intermediate paths that are maybe not needed anymore.

On the con side: * Fuzzing time is lost for the time the fuzzing is stopped, minimized and restarted.

The the big question: * Does a minimized queue/corpus improve finding new coverage or does it hinder it?

The AFL++ team’s own limited analysis seem to to show that keeping intermediate paths help to find more coverage, at least for afl-fuzz.

For honggfuzz in comparison it is a good idea to restart it from time to time if you have other fuzzers (e.g: AFL++) running in parallel to sync the finds of other fuzzers to honggfuzz as it has no syncing feature like AFL++ or libfuzzer.

AFL++ is a great fuzzer if you have the source code available.

However, if there is only the binary program and no source code available, then the standard non-instrumented mode is not effective.

To learn how these binaries can be fuzzed, read fuzzing_binary-only_targets.md.

The short answer is - you cannot, at least not “out of the box”.

For more information on fuzzing network services, see best_practices.md#fuzzing-a-network-service.

Not all GUI programs are suitable for fuzzing. If the GUI program can read the fuzz data from a file without needing any user interaction, then it would be suitable for fuzzing.

For more information on fuzzing GUI programs, see best_practices.md#fuzzing-a-gui-program.

Good performance generally means “making the fuzzing results better”. This can be influenced by various factors, for example, speed (finding lots of paths quickly) or thoroughness (working with decreased speed, but finding better mutations).

There are a few things you can do to improve the fuzzing speed, see best_practices.md#improving-speed.

Stability is measured by how many percent of the edges in the target are “stable”. Sending the same input again and again should take the exact same path through the target every time. If that is the case, the stability is 100%.

If, however, randomness happens, e.g., a thread reading other external data, reaction to timing, etc., then in some of the re-executions with the same data the edge coverage result will be different across runs. Those edges that change are then flagged “unstable”.

The more “unstable” edges there are, the harder it is for AFL++ to identify valid new paths.

If you fuzz in persistent mode (AFL_LOOP or LLVMFuzzerTestOneInput() harnesses, a large number of unstable edges can mean that the target keeps internal state and therefore it is possible that crashes cannot be replayed. In such a case do either not fuzz in persistent mode (remove AFL_LOOP() from your harness or call LLVMFuzzerTestOneInput() harnesses with @@), or set a low AFL_LOOP value, e.g. 100, and enable AFL_PERSISTENT_RECORD in config.h with the same value.

A value above 90% is usually fine and a value above 80% is also still ok, and even a value above 20% can still result in successful finds of bugs. However, it is recommended that for values below 90% or 80% you should take countermeasures to improve stability.

For more information on stability and how to improve the stability value, see best_practices.md#improving-stability.

Not every item in our queue/corpus is the same, some are more interesting, others provide little value. A power schedule measures how “interesting” a value is, and depending on the calculated value spends more or less time mutating it.

AFL++ comes with several power schedules, initially ported from AFLFast, however, modified to be more effective and several more modes added.

The most effective modes are -p fast (default) and -p explore.

If you fuzz with several parallel afl-fuzz instances, then it is beneficial to assign a different schedule to each instance, however the majority should be fast and explore.

It does not make sense to explain the details of the calculation and reasoning behind all of the schedules. If you are interested, read the source code and the AFLFast paper.

It can happen that you see this error on startup when fuzzing a target:

[-] FATAL: forkserver is already up, but an instrumented dlopen() library
           loaded afterwards. You must AFL_PRELOAD such libraries to be able
           to fuzz them or LD_PRELOAD to run outside of afl-fuzz.
           To ignore this set AFL_IGNORE_PROBLEMS=1.

As the error describes, a dlopen() call is happening in the target that is loading an instrumented library after the forkserver is already in place. This is a problem for afl-fuzz because when the forkserver is started, we must know the map size already and it can’t be changed later.

The best solution is to simply set AFL_PRELOAD=foo.so to the libraries that are dlopen’ed (e.g., use strace to see which), or to set a manual forkserver after the final dlopen().

If this is not a viable option, you can set AFL_IGNORE_PROBLEMS=1 but then the existing map will be used also for the newly loaded libraries, which allows it to work, however, the efficiency of the fuzzing will be partially degraded. Note that there is additionally AFL_IGNORE_PROBLEMS_COVERAGE to additionally tell AFL++ to ignore any coverage from the late loaded libaries.

If you see this kind of error when trying to instrument a target with afl-cc/afl-clang-fast/afl-clang-lto:

/prg/tmp/llvm-project/build/bin/clang-13: symbol lookup error: /usr/local/bin/../lib/afl//cmplog-instructions-pass.so: undefined symbol: _ZNK4llvm8TypeSizecvmEv
clang-13: error: unable to execute command: No such file or directory
clang-13: error: clang frontend command failed due to signal (use -v to see invocation)
clang version 13.0.0 (https://github.com/llvm/llvm-project 1d7cf550721c51030144f3cd295c5789d51c4aad)
Target: x86_64-unknown-linux-gnu
Thread model: posix
InstalledDir: /prg/tmp/llvm-project/build/bin
clang-13: note: diagnostic msg:
********************

Then this means that your OS updated the clang installation from an upgrade package and because of that the AFL++ llvm plugins do not match anymore.

Solution: git pull ; make clean install of AFL++.

When you run a large instrumented program stand-alone or via afl-showmap you might see a warning like the following:

Warning: AFL++ tools might need to set AFL_MAP_SIZE to 223723 to be able to run this instrumented program if this crashes!

Depending how the target works it might also crash afterwards.

Solution: just do an export AFL_MAP_SIZE=(the value in the warning).

If you compile C++ harnesses and see undefined reference errors for variables named __afl_..., e.g.:

/usr/bin/ld: /tmp/test-d3085f.o: in function `foo::test()':
test.cpp:(.text._ZN3fooL4testEv[_ZN3fooL4testEv]+0x35): undefined reference to `foo::__afl_connected'
clang: error: linker command failed with exit code 1 (use -v to see invocation)

Then you use AFL++ macros like __AFL_LOOP within a namespace and this will not work.

Solution: Move that harness portion to the global namespace, e.g. before:

#include <cstdio>
namespace foo {
  static void test() {
    while(__AFL_LOOP(1000)) {
      foo::function();
    }
  }
}

int main(int argc, char** argv) {
  foo::test();
  return 0;
}

after:

#include <cstdio>
static void mytest() {
  while(__AFL_LOOP(1000)) {
    foo::function();
  }
}
namespace foo {
  static void test() {
    mytest();
  }
}
int main(int argc, char** argv) {
  foo::test();
  return 0;
}