README.md

metadata

license: mit
task_categories:
  - text-generation
tags:
  - security
  - vulnerability
  - fuzzing
  - oss-fuzz
  - code
  - cyberforge
pretty_name: CyberForge Projects
size_categories:
  - 1K<n<10K

CyberForge Projects

Real OSS-Fuzz C/C++ projects with injected, PoV-validated vulnerabilities, built with CyberForge — 1,034 vulnerabilities across 100 projects.

One archive per project. Each <project>.tar.gz unpacks to that project's tree:

<project>/
├── project.json                        # metadata + secure_base_commit + prebuilt dockerhub_image
├── setup/                              # Dockerfile, build.sh, project.yaml (OSS-Fuzz build)
├── unit_tests/                         # test.sh + parse_results.py
└── vulnerabilities/
    └── vulnerability_<PIPELINE>_<N>/
        ├── inject_vulnerability.diff    # apply -> introduces the vulnerability
        ├── vulnerability_metadata.json  # id, cwe_id, secure_base_commit, ...
        ├── sanitizer_report.txt         # the crash/divergence proving it triggers
        ├── vuln_description.md           # natural-language bug report
        └── exploit_files/
            └── exploit.sh               # proof-of-vulnerability entrypoint

<PIPELINE> is the injection pipeline: FZ (fuzzer-guided), WF_RAG and WF_LLM (agentic).

Usage

# fetch one project and unpack it
hf download AmL-hug/cyberforge-projects guetzli.tar.gz --repo-type dataset --local-dir .
tar xzf guetzli.tar.gz

Key files

• inject_vulnerability.diff — applying it at secure_base_commit introduces the vulnerability; reverting it is the ground-truth fix.
• exploit_files/exploit.sh — runs the proof-of-vulnerability against the built target.
• sanitizer_report.txt — the sanitizer crash (or behavioral divergence) that fires on the vulnerable build and is silent on the secure build.

Related

• Browsable file tree (GitHub mirror): CyberForge-projects
• Framework: CyberForge
• Teacher trajectories: cyberforge-teacher-traj-gemma4-31b and cyberforge-teacher-traj-gpt-5.4-mini