Coverage Guided Fuzz Testing

GitLab allows you to add coverage-guided fuzz testing to your pipelines. This helps you discover bugs and potential security issues that other QA processes may miss. Coverage-guided fuzzing sends random inputs to an instrumented version of your application in an effort to cause unexpected behavior, such as a crash. Such behavior indicates a bug that you should address.

We recommend that you use fuzz testing in addition to the other security scanners in GitLab Secure and your own test processes. If you’re using GitLab CI/CD, you can run your coverage guided fuzz tests as part your CI/CD workflow. You can take advantage of Coverage Guided Fuzzing by including the CI job in your existing .gitlab-ci.yml file.

Supported fuzzing engines and languages

GitLab supports these languages through the fuzzing engine listed for each. We currently provide a Docker image for apps written in Go, but you can test the other languages below by providing a Docker image with the fuzz engine to run your app.

Language Fuzzing Engine Example
C/C++ libFuzzer c-cpp-example
GoLang go-fuzz (libFuzzer support) go-fuzzing-example
Rust cargo-fuzz (libFuzzer support)  

Configuration

To enable fuzzing, you must include the Coverage-Fuzzing.gitlab-ci.yml template provided as part of your GitLab installation.

To do so, add the following to your .gitlab-ci.yml file:

include:
  - template: Coverage-Fuzzing.gitlab-ci.yml

The included template makes available the hidden job .fuzz_base, which you must extend for each of your fuzz targets. Each fuzz target must have a separate job. For example, the go-fuzzing-example project contains one job that extends .fuzz_base for its single fuzz target.

Note that the hidden job .fuzz_base uses several YAML keys that you must not override in your own job. If you include these keys in your own job, you must copy their original content. These keys are:

  • before_script
  • artifacts
  • rules

The my_fuzz_target job (the separate job for your fuzz target) does the following:

  • Extends .fuzz_base.
  • Compiles the fuzz target with go-fuzz.
  • Runs the target with the gitlab-cov-fuzz command, which is available to each job that extends .fuzz_base.
  • Runs on a fuzz stage that usually comes after a test stage.

The gitlab-cov-fuzz is a command-line tool that runs the instrumented application. It parses and analyzes the exception information that the fuzzer outputs. It also downloads the corpus and crash events from previous pipelines automatically. This helps your fuzz targets build on the progress of previous fuzzing jobs. The parsed crash events and data are written to gl-coverage-fuzzing-report.json.

Artifacts

Each fuzzing step outputs these artifacts:

  • gl-coverage-fuzzing-report.json: This file’s format may change in future releases.
  • artifacts.zip: This file contains two directories:
    • corpus: Holds all test cases generated by the current and all previous jobs.
    • crashes: Holds all crash events the current job encountered as well as those not fixed in previous jobs.

Types of Fuzzing Jobs

There are two types of jobs:

  • Fuzzing: Standard fuzzing session. You can configure a long session through a user defined timeout.
  • Regression: Run the fuzz targets through the accumulated test cases generated by previous fuzzing sessions plus fixed crashes from previous sessions. This is usually very quick.

Here’s our current suggestion for configuring your fuzz target’s timeout:

  • Set COVERAGE_FUZZING_BRANCH to the branch where you want to run long-running (async) fuzzing jobs. This is master by default.
  • Use regression or short-running fuzzing jobs for other branches or merge requests.

This suggestion helps find new bugs on the development branch and catch old bugs in merge requests (like unit tests).

You can configure this by passing --regression=false/true to gitlab-cov-fuzz as the Go example shows. Also note that gitlab-cov-fuzz is a wrapper, so you can pass those arguments to configure any option available in the underlying fuzzing engine.

Available variables

Environment variable Description
COVERAGE_FUZZING_BRANCH The branch for long-running fuzzing jobs. The default is master.
CI_SEED_CORPUS Path to a seed corpus directory. The default is empty.

The files in the seed corpus (CI_SEED_CORPUS), if provided, aren’t updated unless you commit new files to your Git repository. There’s usually no need to frequently update the seed corpus. As part of the GitLab artifacts system, GitLab saves in a corpus directory the new test cases that every run generates. In any subsequent runs, GitLab also reuses the generated corpus together with the seed corpus.

Reports JSON format

The gitlab-cov-fuzz tool emits a JSON report file. For more information, see the schema for this report.

You can download the JSON report file from the CI pipelines page. For more information, see Downloading artifacts.

Here’s an example Coverage Fuzzing report:

{
  "version": "v1.0.8",
  "regression": false,
  "exit_code": -1,
  "vulnerabilities": [
    {
      "category": "coverage_fuzzing",
      "message": "Heap-buffer-overflow\nREAD 1",
      "description": "Heap-buffer-overflow\nREAD 1",
      "severity": "Critical",
      "stacktrace_snippet": "INFO: Seed: 3415817494\nINFO: Loaded 1 modules   (7 inline 8-bit counters): 7 [0x10eee2470, 0x10eee2477), \nINFO: Loaded 1 PC tables (7 PCs): 7 [0x10eee2478,0x10eee24e8), \nINFO:        5 files found in corpus\nINFO: -max_len is not provided; libFuzzer will not generate inputs larger than 4096 bytes\nINFO: seed corpus: files: 5 min: 1b max: 4b total: 14b rss: 26Mb\n#6\tINITED cov: 7 ft: 7 corp: 5/14b exec/s: 0 rss: 26Mb\n=================================================================\n==43405==ERROR: AddressSanitizer: heap-buffer-overflow on address 0x602000001573 at pc 0x00010eea205a bp 0x7ffee0d5e090 sp 0x7ffee0d5e088\nREAD of size 1 at 0x602000001573 thread T0\n    #0 0x10eea2059 in FuzzMe(unsigned char const*, unsigned long) fuzz_me.cc:9\n    #1 0x10eea20ba in LLVMFuzzerTestOneInput fuzz_me.cc:13\n    #2 0x10eebe020 in fuzzer::Fuzzer::ExecuteCallback(unsigned char const*, unsigned long) FuzzerLoop.cpp:556\n    #3 0x10eebd765 in fuzzer::Fuzzer::RunOne(unsigned char const*, unsigned long, bool, fuzzer::InputInfo*, bool*) FuzzerLoop.cpp:470\n    #4 0x10eebf966 in fuzzer::Fuzzer::MutateAndTestOne() FuzzerLoop.cpp:698\n    #5 0x10eec0665 in fuzzer::Fuzzer::Loop(std::__1::vector\u003cfuzzer::SizedFile, fuzzer::fuzzer_allocator\u003cfuzzer::SizedFile\u003e \u003e\u0026) FuzzerLoop.cpp:830\n    #6 0x10eead0cd in fuzzer::FuzzerDriver(int*, char***, int (*)(unsigned char const*, unsigned long)) FuzzerDriver.cpp:829\n    #7 0x10eedaf82 in main FuzzerMain.cpp:19\n    #8 0x7fff684fecc8 in start+0x0 (libdyld.dylib:x86_64+0x1acc8)\n\n0x602000001573 is located 0 bytes to the right of 3-byte region [0x602000001570,0x602000001573)\nallocated by thread T0 here:\n    #0 0x10ef92cfd in wrap__Znam+0x7d (libclang_rt.asan_osx_dynamic.dylib:x86_64+0x50cfd)\n    #1 0x10eebdf31 in fuzzer::Fuzzer::ExecuteCallback(unsigned char const*, unsigned long) FuzzerLoop.cpp:541\n    #2 0x10eebd765 in fuzzer::Fuzzer::RunOne(unsigned char const*, unsigned long, bool, fuzzer::InputInfo*, bool*) FuzzerLoop.cpp:470\n    #3 0x10eebf966 in fuzzer::Fuzzer::MutateAndTestOne() FuzzerLoop.cpp:698\n    #4 0x10eec0665 in fuzzer::Fuzzer::Loop(std::__1::vector\u003cfuzzer::SizedFile, fuzzer::fuzzer_allocator\u003cfuzzer::SizedFile\u003e \u003e\u0026) FuzzerLoop.cpp:830\n    #5 0x10eead0cd in fuzzer::FuzzerDriver(int*, char***, int (*)(unsigned char const*, unsigned long)) FuzzerDriver.cpp:829\n    #6 0x10eedaf82 in main FuzzerMain.cpp:19\n    #7 0x7fff684fecc8 in start+0x0 (libdyld.dylib:x86_64+0x1acc8)\n\nSUMMARY: AddressSanitizer: heap-buffer-overflow fuzz_me.cc:9 in FuzzMe(unsigned char const*, unsigned long)\nShadow bytes around the buggy address:\n  0x1c0400000250: fa fa fd fa fa fa fd fa fa fa fd fa fa fa fd fa\n  0x1c0400000260: fa fa fd fa fa fa fd fa fa fa fd fa fa fa fd fa\n  0x1c0400000270: fa fa fd fa fa fa fd fa fa fa fd fa fa fa fd fa\n  0x1c0400000280: fa fa fd fa fa fa fd fa fa fa fd fa fa fa fd fa\n  0x1c0400000290: fa fa fd fa fa fa fd fa fa fa fd fa fa fa fd fa\n=\u003e0x1c04000002a0: fa fa fd fa fa fa fd fa fa fa fd fa fa fa[03]fa\n  0x1c04000002b0: fa fa fa fa fa fa fa fa fa fa fa fa fa fa fa fa\n  0x1c04000002c0: fa fa fa fa fa fa fa fa fa fa fa fa fa fa fa fa\n  0x1c04000002d0: fa fa fa fa fa fa fa fa fa fa fa fa fa fa fa fa\n  0x1c04000002e0: fa fa fa fa fa fa fa fa fa fa fa fa fa fa fa fa\n  0x1c04000002f0: fa fa fa fa fa fa fa fa fa fa fa fa fa fa fa fa\nShadow byte legend (one shadow byte represents 8 application bytes):\n  Addressable:           00\n  Partially addressable: 01 02 03 04 05 06 07 \n  Heap left redzone:       fa\n  Freed heap region:       fd\n  Stack left redzone:      f1\n  Stack mid redzone:       f2\n  Stack right redzone:     f3\n  Stack after return:      f5\n  Stack use after scope:   f8\n  Global redzone:          f9\n  Global init order:       f6\n  Poisoned by user:        f7\n  Container overflow:      fc\n  Array cookie:            ac\n  Intra object redzone:    bb\n  ASan internal:           fe\n  Left alloca redzone:     ca\n  Right alloca redzone:    cb\n  Shadow gap:              cc\n==43405==ABORTING\nMS: 1 EraseBytes-; base unit: de3a753d4f1def197604865d76dba888d6aefc71\n0x46,0x55,0x5a,\nFUZ\nartifact_prefix='./crashes/'; Test unit written to ./crashes/crash-0eb8e4ed029b774d80f2b66408203801cb982a60\nBase64: RlVa\nstat::number_of_executed_units: 122\nstat::average_exec_per_sec:     0\nstat::new_units_added:          0\nstat::slowest_unit_time_sec:    0\nstat::peak_rss_mb:              28",
      "scanner": {
        "id": "libFuzzer",
        "name": "libFuzzer"
      },
      "location": {
        "crash_address": "0x602000001573",
        "crash_state": "FuzzMe\nstart\nstart+0x0\n\n",
        "crash_type": "Heap-buffer-overflow\nREAD 1"
      },
      "tool": "libFuzzer"
    }
  ]
}

Additional Configuration

The gitlab-cov-fuzz command passes all arguments it receives to the underlying fuzzing engine. You can therefore use all the options available in that fuzzing engine. For more information on these options, see the underlying fuzzing engine’s documentation.

Glossary

  • Seed corpus: The set of test cases given as initial input to the fuzz target. This usually speeds up the fuzz target substantially. This can be either manually created test cases or auto-generated with the fuzz target itself from previous runs.
  • Corpus: The set of meaningful test cases that are generated while the fuzzer is running. Each meaningful test case produces new coverage in the tested program. It’s advised to re-use the corpus and pass it to subsequent runs.