path: root/docs/custom_mutators.md

# Custom Mutators in AFL++

This file describes how you can implement custom mutations to be used in AFL.
For now, we support C/C++ library and Python module, collectivelly named as the
custom mutator.

There is also experimental support for Rust in `custom_mutators/rust`.
Please refer to that directory for documentation.
Run ```cargo doc -p custom_mutator --open``` in that directory to view the
documentation in your web browser.

Implemented by
- C/C++ library (`*.so`): Khaled Yakdan from Code Intelligence (<yakdan@code-intelligence.de>)
- Python module: Christian Holler from Mozilla (<choller@mozilla.com>)

## 1) Introduction

Custom mutators can be passed to `afl-fuzz` to perform custom mutations on test
cases beyond those available in AFL. For example, to enable structure-aware
fuzzing by using libraries that perform mutations according to a given grammar.

The custom mutator is passed to `afl-fuzz` via the `AFL_CUSTOM_MUTATOR_LIBRARY`
or `AFL_PYTHON_MODULE` environment variable, and must export a fuzz function.
Now afl also supports multiple custom mutators which can be specified in the same `AFL_CUSTOM_MUTATOR_LIBRARY` environment variable like this.
```bash
export AFL_CUSTOM_MUTATOR_LIBRARY="full/path/to/mutator_first.so;full/path/to/mutator_second.so"
```
Please see [APIs](#2-apis) and [Usage](#3-usage) for detail.

The custom mutation stage is set to be the first non-deterministic stage (right before the havoc stage).

Note: If `AFL_CUSTOM_MUTATOR_ONLY` is set, all mutations will solely be
performed with the custom mutator.

## 2) APIs

C/C++:
```c
void *afl_custom_init(afl_state_t *afl, unsigned int seed);
unsigned int afl_custom_fuzz_count(void *data, const unsigned char *buf, size_t buf_size);
size_t afl_custom_fuzz(void *data, unsigned char *buf, size_t buf_size, unsigned char **out_buf, unsigned char *add_buf, size_t add_buf_size, size_t max_size);
const char *afl_custom_describe(void *data, size_t max_description_len);
size_t afl_custom_post_process(void *data, unsigned char *buf, size_t buf_size, unsigned char **out_buf);
int afl_custom_init_trim(void *data, unsigned char *buf, size_t buf_size);
size_t afl_custom_trim(void *data, unsigned char **out_buf);
int afl_custom_post_trim(void *data, unsigned char success);
size_t afl_custom_havoc_mutation(void *data, unsigned char *buf, size_t buf_size, unsigned char **out_buf, size_t max_size);
unsigned char afl_custom_havoc_mutation_probability(void *data);
unsigned char afl_custom_queue_get(void *data, const unsigned char *filename);
void afl_custom_queue_new_entry(void *data, const unsigned char *filename_new_queue, const unsigned int *filename_orig_queue);
const char* afl_custom_introspection(my_mutator_t *data);
void afl_custom_deinit(void *data);
```

Python:
```python
def init(seed):
    pass

def fuzz_count(buf, add_buf, max_size):
    return cnt

def fuzz(buf, add_buf, max_size):
    return mutated_out

def describe(max_description_length):
    return "description_of_current_mutation"

def post_process(buf):
    return out_buf

def init_trim(buf):
    return cnt

def trim():
    return out_buf

def post_trim(success):
    return next_index

def havoc_mutation(buf, max_size):
    return mutated_out

def havoc_mutation_probability():
    return probability # int in [0, 100]

def queue_get(filename):
    return True

def queue_new_entry(filename_new_queue, filename_orig_queue):
    pass

def introspection():
    return string

def deinit():  # optional for Python
    pass
```

### Custom Mutation

- `init`:

    This method is called when AFL++ starts up and is used to seed RNG and set up buffers and state.

- `queue_get` (optional):

    This method determines whether the custom fuzzer should fuzz the current
    queue entry or not

- `fuzz_count` (optional):

    When a queue entry is selected to be fuzzed, afl-fuzz selects the number
    of fuzzing attempts with this input based on a few factors.
    If however the custom mutator wants to set this number instead on how often
    it is called for a specific queue entry, use this function.
    This function is most useful if `AFL_CUSTOM_MUTATOR_ONLY` is **not** used.

- `fuzz` (optional):

    This method performs custom mutations on a given input. It also accepts an
    additional test case.
    Note that this function is optional - but it makes sense to use it.
    You would only skip this if `post_process` is used to fix checksums etc.
    so if you are using it e.g. as a post processing library.
    Note that a length > 0 *must* be returned!

- `describe` (optional):

    When this function is called, it shall describe the current testcase,
    generated by the last mutation. This will be called, for example,
    to name the written testcase file after a crash occurred.
    Using it can help to reproduce crashing mutations.

- `havoc_mutation` and `havoc_mutation_probability` (optional):

    `havoc_mutation` performs a single custom mutation on a given input. This
    mutation is stacked with other mutations in havoc. The other method,
    `havoc_mutation_probability`, returns the probability that `havoc_mutation`
    is called in havoc. By default, it is 6%.

- `post_process` (optional):

    For some cases, the format of the mutated data returned from the custom
    mutator is not suitable to directly execute the target with this input.
    For example, when using libprotobuf-mutator, the data returned is in a
    protobuf format which corresponds to a given grammar. In order to execute
    the target, the protobuf data must be converted to the plain-text format
    expected by the target. In such scenarios, the user can define the
    `post_process` function. This function is then transforming the data into the
    format expected by the API before executing the target.

    This can return any python object that implements the buffer protocol and
    supports PyBUF_SIMPLE. These include bytes, bytearray, etc.

- `queue_new_entry` (optional):

    This methods is called after adding a new test case to the queue.

- `introspection` (optional):

    This method is called after a new queue entry, crash or timeout is
    discovered if compiled with INTROSPECTION. The custom mutator can then
    return a string (const char *) that reports the exact mutations used.

- `deinit`:

    The last method to be called, deinitializing the state.

Note that there are also three functions for trimming as described in the
next section.

### Trimming Support

The generic trimming routines implemented in AFL++ can easily destroy the
structure of complex formats, possibly leading to a point where you have a lot
of test cases in the queue that your Python module cannot process anymore but
your target application still accepts. This is especially the case when your
target can process a part of the input (causing coverage) and then errors out
on the remaining input.

In such cases, it makes sense to implement a custom trimming routine. The API
consists of multiple methods because after each trimming step, we have to go
back into the C code to check if the coverage bitmap is still the same for the
trimmed input. Here's a quick API description:

- `init_trim` (optional):

    This method is called at the start of each trimming operation and receives
    the initial buffer. It should return the amount of iteration steps possible
    on this input (e.g. if your input has n elements and you want to remove them
    one by one, return n, if you do a binary search, return log(n), and so on).

    If your trimming algorithm doesn't allow to determine the amount of
    (remaining) steps easily (esp. while running), then you can alternatively
    return 1 here and always return 0 in `post_trim` until you are finished and
    no steps remain. In that case, returning 1 in `post_trim` will end the
    trimming routine. The whole current index/max iterations stuff is only used
    to show progress.

- `trim` (optional)

    This method is called for each trimming operation. It doesn't have any
    arguments because we already have the initial buffer from `init_trim` and we
    can memorize the current state in the data variables. This can also save
    reparsing steps for each iteration. It should return the trimmed input
    buffer, where the returned data must not exceed the initial input data in
    length. Returning anything that is larger than the original data (passed to
    `init_trim`) will result in a fatal abort of AFL++.

- `post_trim` (optional)

    This method is called after each trim operation to inform you if your
    trimming step was successful or not (in terms of coverage). If you receive
    a failure here, you should reset your input to the last known good state.
    In any case, this method must return the next trim iteration index (from 0
    to the maximum amount of steps you returned in `init_trim`).

Omitting any of three trimming methods will cause the trimming to be disabled
and trigger a fallback to the builtin default trimming routine.

### Environment Variables

Optionally, the following environment variables are supported:

- `AFL_CUSTOM_MUTATOR_ONLY`

    Disable all other mutation stages. This can prevent broken testcases
    (those that your Python module can't work with anymore) to fill up your
    queue. Best combined with a custom trimming routine (see below) because
    trimming can cause the same test breakage like havoc and splice.


- `AFL_PYTHON_ONLY`

    Deprecated and removed, use `AFL_CUSTOM_MUTATOR_ONLY` instead.

- `AFL_DEBUG`

    When combined with `AFL_NO_UI`, this causes the C trimming code to emit
    additional messages about the performance and actions of your custom
    trimmer. Use this to see if it works :)

## 3) Usage

### Prerequisite

For Python mutators, the python 3 or 2 development package is required. On
Debian/Ubuntu/Kali it can be installed like this:

```bash
sudo apt install python3-dev
# or
sudo apt install python-dev
```

Then, AFL++ can be compiled with Python support. The AFL++ Makefile detects
Python 2 and 3 through `python-config` if it is in the PATH and compiles
`afl-fuzz` with the feature if available.

Note: for some distributions, you might also need the package `python[23]-apt`.
In case your setup is different, set the necessary variables like this:
`PYTHON_INCLUDE=/path/to/python/include LDFLAGS=-L/path/to/python/lib make`.

### Custom Mutator Preparation

For C/C++ mutators, the source code must be compiled as a shared object:
```bash
gcc -shared -Wall -O3 example.c -o example.so
```
Note that if you specify multiple custom mutators, the corresponding functions will
be called in the order in which they are specified. e.g first `post_process` function of
`example_first.so` will be called and then that of `example_second.so`.

### Run

C/C++
```bash
export AFL_CUSTOM_MUTATOR_LIBRARY="/full/path/to/example_first.so;/full/path/to/example_second.so"
afl-fuzz /path/to/program
```

Python
```bash
export PYTHONPATH=`dirname /full/path/to/example.py`
export AFL_PYTHON_MODULE=example
afl-fuzz /path/to/program
```

## 4) Example

Please see [example.c](../custom_mutators/examples/example.c) and
[example.py](../custom_mutators/examples/example.py)

## 5) Other Resources

- AFL libprotobuf mutator
    - [bruce30262/libprotobuf-mutator_fuzzing_learning](https://github.com/bruce30262/libprotobuf-mutator_fuzzing_learning/tree/master/4_libprotobuf_aflpp_custom_mutator)
    - [thebabush/afl-libprotobuf-mutator](https://github.com/thebabush/afl-libprotobuf-mutator)
- [XML Fuzzing@NullCon 2017](https://www.agarri.fr/docs/XML_Fuzzing-NullCon2017-PUBLIC.pdf)
    - [A bug detected by AFL + XML-aware mutators](https://bugs.chromium.org/p/chromium/issues/detail?id=930663)