Unicorn-based binary-only instrumentation for afl-fuzz

The idea and much of the original implementation comes from Nathan Voss njvoss299@gmail.com.

The port to AFL++ is by Dominik Maier mail@dmnk.co.

The CompareCoverage and NeverZero counters features are by Andrea Fioraldi andreafioraldi@gmail.com.

1) Introduction

The code in unicorn_mode/ allows you to build the Unicorn Engine with AFL++ support. This means, you can run anything that can be emulated in unicorn and obtain instrumentation output for black-box, closed-source binary code snippets. This mechanism can be then used by afl-fuzz to stress-test targets that couldn't be built with afl-cc or used in QEMU mode.

There is a significant performance penalty compared to native AFL, but at least we're able to use AFL++ on these binaries, right?

2) How to use

First, you will need a working harness for your target in unicorn, using Python, C, or Rust.

For some pointers for more advanced emulation, take a look at BaseSAFE and Qiling.

Building AFL++'s Unicorn mode

First, make AFL++ as usual. Once that completes successfully, you need to build and add in the Unicorn mode features:

cd unicorn_mode
./build_unicorn_support.sh

NOTE: This script checks out a Unicorn Engine fork as submodule that has been tested and is stable-ish, based on the unicorn engine next branch.

Building Unicorn will take a little bit (~5-10 minutes). Once it completes, it automatically compiles a sample application and verifies that it works.

Fuzzing with Unicorn mode

To use unicorn-mode effectively, you need to prepare the following:

Relevant binary code to be fuzzed
Knowledge of the memory map and good starting state
Folder containing sample inputs to start fuzzing with
- Same ideas as any other AFL++ inputs
- Quality/speed of results will depend greatly on the quality of starting samples
- See AFL's guidance on how to create a sample corpus
Unicornafl-based test harness in Rust, C, or Python, which:
- Adds memory map regions
- Loads binary code into memory
- Calls uc.afl_fuzz() / uc.afl_start_forkserver
- Loads and verifies data to fuzz from a command-line specified file
  - AFL++ will provide mutated inputs by changing the file passed to the test harness
  - Presumably the data to be fuzzed is at a fixed buffer address
  - If input constraints (size, invalid bytes, etc.) are known, they should be checked in the place_input handler. If a constraint fails, just return false from the handler. AFL++ will treat the input as 'uninteresting' and move on.
- Sets up registers and memory state to start testing
- Emulates the interesting code from beginning to end
- If a crash is detected, the test harness must 'crash' by throwing a signal (SIGSEGV, SIGKILL, SIGABORT, etc.), or indicate a crash in the crash validation callback.

Once you have all those things ready to go, you just need to run afl-fuzz in unicorn-mode by passing in the -U flag:

afl-fuzz -U -m none -i /path/to/inputs -o /path/to/results -- ./test_harness @@

The normal afl-fuzz command line format applies to everything here. Refer to AFL's main documentation for more info about how to use afl-fuzz effectively.

For a much clearer vision of what all of this looks like, refer to the sample provided in the samples/ directory. There is also a blog post that uses slightly older concepts, but describes the general ideas, at:

https://medium.com/@njvoss299/afl-unicorn-fuzzing-arbitrary-binary-code-563ca28936bf

The helper_scripts/ directory also contains several helper scripts that allow you to dump context from a running process, load it, and hook heap allocations. For details on how to use this, check out the follow-up blog post to the one linked above:

https://hackernoon.com/afl-unicorn-part-2-fuzzing-the-unfuzzable-bea8de3540a5

An example use of AFL-Unicorn mode is discussed in the paper Unicorefuzz: https://www.usenix.org/conference/woot19/presentation/maier

3) Options

As for the QEMU-based instrumentation, unicornafl comes with a sub-instruction based instrumentation similar in purpose to laf-intel.

The options that enable Unicorn CompareCoverage are the same used for QEMU. This will split up each multi-byte compare to give feedback for each correct byte:

AFL_COMPCOV_LEVEL=1 to instrument comparisons with only immediate values.
AFL_COMPCOV_LEVEL=2 to instrument all comparison instructions.

Comparison instructions are currently instrumented only for the x86, x86_64, and ARM targets.

4) Gotchas, feedback, bugs

Running the build script builds unicornafl and its Python bindings and installs them on your system. This installation will leave any existing Unicorn installations untouched.

If you want to use unicornafl instead of unicorn in a script, replace all unicorn imports with unicornafl inputs, everything else should "just work". If you use 3rd party code depending on unicorn, you can use unicornafl monkeypatching. Before importing anything that depends on unicorn, do:

import unicornafl
unicornafl.monkeypatch()

This will replace all unicorn imports with unicornafl inputs.

5) Examples

Apart from reading the documentation in afl.c and the Python bindings of unicornafl, the best documentation are the samples/.

The following examples exist at the time of writing:

c: A simple example on how to use the C bindings
compcov_x64: A Python example that uses compcov to traverse hard-to-reach blocks
persistent: A C example using persistent mode for maximum speed, and resetting the target state between each iteration
python_simple: A simple Python example
speedtest/c: The C harness for an example target, used to compare C, Python, and Rust bindings and fix speed issues
speedtest/python: Fuzzing the same target in Python
speedtest/rust: Fuzzing the same target using a Rust harness

Usually, the place to look at is the harness in each folder. The source code in each harness is pretty well documented. Most harnesses also have the afl-fuzz commandline, or even offer a make fuzz Makefile target. Targets in these folders, if x86, can usually be made using make target in each folder or get shipped pre-built (plus their source).

Especially take a look at the speedtest documentation to see how the languages compare.

Uh oh!