- Creating the file structure
- Disk space restrictions
- Fuzzer execution environment
- Testing locally
- Debugging Problems
- Efficient fuzzing
- Checking in to the OSS-Fuzz repository
- Reviewing results
Before you can start setting up your new project for fuzzing, you must do the following:
If you want to run
sudo, you can create a docker group.
Note: Docker images can consume significant disk space. Run docker-cleanup periodically to garbage-collect unused images.
Each project directory also contains the following three configuration files:
- project.yaml - provides metadata about the project.
- Dockerfile - defines the container environment with information on dependencies needed to build the project and its fuzz targets.
- build.sh - defines the build script that executes inside the Docker container and generates the project build.
You can automatically create a new directory for your project in OSS-Fuzz and generate templated versions of the configuration files by running the following commands:
$ cd /path/to/oss-fuzz $ export PROJECT_NAME=<project_name> $ python infra/helper.py generate $PROJECT_NAME
Once the template configuration files are created, you can modify them to fit your project.
Note: We prefer that you keep and maintain fuzz targets in your own source code repository. If this isn’t possible, you can store them inside the OSS-Fuzz project directory you created.
This configuration file stores project metadata. The following attributes are supported:
- vendor_ccs (optional)
- sanitizers (optional)
- architectures (optional)
- help_url (optional)
- builds_per_day (optional)
You project’s homepage.
Programming language the project is written in. Values you can specify include:
The primary contact and list of other contacts to be CCed. Each person listed gets access to ClusterFuzz, including crash reports and fuzzer statistics, and are auto-cced on new bugs filed in the OSS-Fuzz tracker. If you’re a primary or a CC, you’ll need to use a Google account to get full access. (why?).
Path to source code repository hosting the code, e.g.
The list of vendor email addresses that are downstream consumers of the project and want access to the bug reports as they are filed.
Any changes to this list must follow these rules:
- Approved by the project maintainer (e.g. comment on pull request, reply on project mailing list).
- An organization email address is used.
MemorySanitizer (“memory”) is also supported and recommended, but is not enabled by default due to the likelihood of false positives from un-instrumented system dependencies. If you want to use “memory,” first make sure your project’s runtime dependencies are listed in the OSS-Fuzz msan-libs-builder Dockerfile. Then, you can opt in by adding “memory” to your list of sanitizers.
If your project does not build with a particular sanitizer configuration and you need some time to fix it, you can use
sanitizers to override the defaults temporarily. For example, to disable the UndefinedBehaviourSanitizer build, just specify all supported sanitizers except “undefined”.
If you want to test a particular sanitizer to see what crashes it generates without filing them in the issue tracker, you can set an
experimental flag. For example, if you want to test “memory”, set
experimental: True like this:
sanitizers: - address - memory: experimental: True - undefined
Crashes can be accessed on the ClusterFuzz homepage.
sanitizers example: boringssl.
The list of architectures to fuzz on. ClusterFuzz supports fuzzing on x86_64 (aka x64) by default. However you can also fuzz using AddressSanitizer and libFuzzer on i386 (aka x86, or 32 bit) by specifying “x86_64” and “i386” in “architectures” like this:
architectures: - x86_64 - i386
By fuzzing on i386 you might find bugs that:
- Only occur in architecture-specific source code (e.g. code that contains i386 assembly).
- Exist in architecture-independent source code and which only affects i386 users.
- Exist in architecture-independent source code and which affects users on other 32-bit platforms such as AArch32 (aka 32-bit ARM).
Note that some bugs which affect x86_64 may be discovered on i386 and filed as such. On the testcase page of each oss-fuzz issue is a list of other jobs where the crash reproduces, this can let you know if the crash exists on x86_64 as well.
Fuzzing on i386 is not enabled by default because many projects won’t build for i386 without some modification to their OSS-Fuzz build process. For example, you will need to link against
$LIB_FUZZING_ENGINE and possibly install i386 dependencies within the x86_64 docker image (for example) to get things working.
A link to a custom help URL that appears in bug reports instead of the default OSS-Fuzz guide to reproducing crashes. This can be useful if you assign bugs to members of your project unfamiliar with OSS-Fuzz, or if they should follow a different workflow for reproducing and fixing bugs than the standard one outlined in the reproducing guide.
help_url example: skia.
The number of times the project should be built per day. OSS-Fuzz allows upto 4 builds per day, and builds once per day by default. Example:
Will build the project twice per day.
This configuration file defines the Docker image for your project. Your build.sh script will be executed in inside the container you define. For most projects, the image is simple:
FROM gcr.io/oss-fuzz-base/base-builder # base image with clang toolchain RUN apt-get update && apt-get install -y ... # install required packages to build your project RUN go get ... # install dependencies to build your Go project RUN git clone <git_url> <checkout_dir> # checkout all sources needed to build your project WORKDIR <checkout_dir> # current directory for the build script COPY build.sh fuzzer.cc $SRC/ # copy build script and other fuzzer files in src dir
In the above example, the git clone will check out the source to
In general, this script should do the following:
- Build the project using your build system with the correct compiler.
- Provide compiler flags as environment variables.
- Build your fuzz targets and link your project’s build with libFuzzer.
Resulting binaries should be placed in
Here’s an example from Expat (source):
#!/bin/bash -eu ./buildconf.sh # configure scripts usually use correct environment variables. ./configure make clean make -j$(nproc) all $CXX $CXXFLAGS -std=c++11 -Ilib/ \ $SRC/parse_fuzzer.cc -o $OUT/parse_fuzzer \ $LIB_FUZZING_ENGINE .libs/libexpat.a cp $SRC/*.dict $SRC/*.options $OUT/
If your project is written in Go, check out the Integrating a Go project page.
- Don’t assume the fuzzing engine is libFuzzer by default, because we generate builds for libFuzzer, AFL and Honggfuzz fuzzing engine configurations. Instead, link the fuzzing engine using $LIB_FUZZING_ENGINE.
- Make sure that the binary names for your fuzz targets contain only alphanumeric characters, underscore(_) or dash(-). Otherwise, they won’t run on our infrastructure.
- Don’t remove source code files. They are needed for code coverage.
When your build.sh script is executed, the following locations are available within the image:
| || ||Directory to store build artifacts (fuzz targets, dictionaries, options files, seed corpus archives).|
| || ||Directory to checkout source files.|
| || ||Directory to store intermediate files.|
Although the files layout is fixed within a container, environment variables are provided so you can write retargetable scripts.
In case your fuzz target uses the FuzzedDataProvider class, make sure it is included via
#include <fuzzer/FuzzedDataProvider.h> directive.
Only binaries without an extension are accepted as targets. Extensions are reserved for other artifacts, like .dict.
You must use the special compiler flags needed to build your project and fuzz targets. These flags are provided in the following environment variables:
| ||The C and C++ compiler binaries.|
| ||C and C++ compiler flags.|
| ||C++ compiler argument to link fuzz target against the prebuilt engine library (e.g. libFuzzer).|
You must use
$CXX as a linker, even if your project is written in pure C.
Most well-crafted build scripts will automatically use these variables. If not, pass them manually to the build tool.
See the Provided Environment Variables section in
base-builder image documentation for more details.
Our builders have a disk size of 250GB (this includes space taken up by the OS). Builds must keep peak disk usage below this.
In addition, please keep the size of the build (everything copied to
$OUT) small (<10GB uncompressed). The build is repeatedly transferred and unzipped during fuzzing and runs on VMs with limited disk space.
You can build your docker image and fuzz targets locally, so you can test them before you push them to the OSS-Fuzz repository.
Run the same helper script you used to create your directory structure, this time using it to build your docker image and fuzz targets:
$ cd /path/to/oss-fuzz $ python infra/helper.py build_image $PROJECT_NAME $ python infra/helper.py build_fuzzers --sanitizer <address/memory/undefined> $PROJECT_NAME
The built binaries appear in the
/path/to/oss-fuzz/build/out/$PROJECT_NAMEdirectory on your machine (and
$OUTin the container).
Note: You must run your fuzz target binaries inside the base-runner docker container to make sure that they work properly.
Find failures to fix by running the
$ python infra/helper.py check_build $PROJECT_NAME
If you want to test changes against a particular fuzz target, run the following command:
$ python infra/helper.py run_fuzzer $PROJECT_NAME <fuzz_target> --corpus-dir=<path-to-temp-corpus-dir>
We recommend taking a look at your code coverage as a sanity check to make sure that your fuzz targets get to the code you expect. This would use the corpus generated from the previous
run_fuzzerstep in your local corpus directory.
$ python infra/helper.py build_fuzzers --sanitizer coverage $PROJECT_NAME $ python infra/helper.py coverage $PROJECT_NAME --fuzz-target=<fuzz_target> --corpus-dir=<path-to-temp-corpus-dir>
Please refer to code coverage for detailed information on code coverage generation.
Note: Currently, we only support AddressSanitizer (address) and UndefinedBehaviorSanitizer (undefined) configurations. MemorySanitizer is recommended, but needs to be enabled manually once you verify that all system dependencies are instrumented. Make sure to test each of the supported build configurations with the above commands (build_fuzzers -> run_fuzzer -> coverage).
If everything works locally, it should also work on our automated builders and ClusterFuzz. If you check in your files and experience failures, review your dependencies.
To improve your fuzz target ability to find bugs faster, you should consider the following ways:
OSS-Fuzz uses evolutionary fuzzing algorithms. Supplying seed corpus consisting of good sample inputs is one of the best ways to improve fuzz target’s coverage.
To provide a corpus for
my_fuzzer_seed_corpus.zip file next to the fuzz target’s binary in
$OUT during the build. Individual files in this archive will be used as starting inputs for mutations. The name of each file in the corpus is the sha1 checksum (which you can get using the
shasum comand) of its contents. You can store the corpus next to source files, generate during build or fetch it using curl or any other tool of your choice. (example: boringssl).
Seed corpus files will be used for cross-mutations and portions of them might appear in bug reports or be used for further security research. It is important that corpus has an appropriate and consistent license.
See also Accessing Corpora for information about getting access to the corpus we are currently using for your fuzz targets.
Dictionaries hugely improve fuzzing efficiency for inputs with lots of similar sequences of bytes. libFuzzer documentation
Put your dict file in
$OUT. If the dict filename is the same as your target binary name (i.e.
%fuzz_target%.dict), it will be automatically used. If the name is different (e.g. because it is shared by several targets), specify this in .options file:
[libfuzzer] dict = dictionary_name.dict
By default, the fuzzing engine will generate input of any arbitrary length. This might be useful to try corner cases that could lead to a security vulnerability. However, if large inputs are not necessary to increase the coverage of your target API, it is important to add a limit here to significantly improve performance.
if (size < kMinInputLength || size > kMaxInputLength) return 0;
Once you’ve tested your fuzzing files locally, fork OSS-Fuzz, commit, and push to the fork. Then create a pull request with your change. Follow the Forking Project guide if you’re new to contributing via GitHub.
Please include copyright headers for all files checked in to oss-fuzz:
# Copyright 2020 Google LLC # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. # ################################################################################
Exception: If you’re porting a fuzz target from Chromium, keep the original Chromium license header.
Once your change is merged, your project and fuzz targets should be automatically built and run on ClusterFuzz after a short while (< 1 day). If you think there’s a problem, you can check your project’s build status.
Use the ClusterFuzz web interface to review the following:
- Crashes generated
- Code coverage statistics
- Fuzzer statistics
- Fuzzer performance analyzer (linked from fuzzer statistics)
Note: Your Google Account must be listed in project.yaml for you to have access to the ClusterFuzz web interface.
Once your project has started building, we’d love it if you added our badge in your project’s README. This allows you to see bugs found by your OSS-Fuzz integration at a glance. See brotli’s README for an example.
Adding it is super easy, just follow this template: