exploitgym

ExploitGym

ExploitGym is a large-scale, realistic benchmark built from real-world vulnerabilities across userspace programs, Google's V8 engine, and the Linux kernel, designed to evaluate AI agents' ability to develop exploits.

Quick start

…

No secret is baked into the repo: the controller generates its token salt, flag seed, and API key on startup (or takes them from those env vars) and both ends must use the same values — see Controller secrets.

Detailed setup steps (system dependencies, GDB, static node, agent CLIs) live in docs/setup.md.

Documentation

• Setup: Python deps, GDB, socat/nc, node + agent CLIs
• Docker images: pulling target images per task family
• Evaluation: controller / firewall / LLM proxy + examples/run_agent.py
• Defenses: disabling system defenses (ASLR, etc.)
• Firewall: outbound network isolation for agent containers
• Submission: submission format and requirements for the benchmark leaderboard

Benchmark updates

The released benchmark is actively maintained. The current release is v1.0 with 869 instances. See CHANGELOG.md for the full version history. The canonical task list for the current release is data/task_ids/v1.txt.

Citation

If you use ExploitGym in your research, please cite:

@article{wang2026exploitgym,
  title={ExploitGym: Can AI Agents Turn Security Vulnerabilities into Real Attacks?},
  author={Wang, Zhun and Schiller, Nico and Li, Hongwei and Sesha Narayana, Srijiith and Nasr, Milad and Carlini, Nicholas and Qi, Xiangyu and Wallace, Eric and Bursztein, Elie and Invernizzi, Luca and Thomas, Kurt and Shoshitaishvili, Yan and Guo, Wenbo and He, Jingxuan and Holz, Thorsten and Song, Dawn},
  journal={arXiv preprint arXiv:2605.11086},
  year={2026}
}

License

The source code is licensed under Apache-2.0. The bundled task data under data/tasks/ derives from external upstreams and retains their respective licenses, see DATA_LICENSE.md.