shirman/exploitgym-answers

ExploitGym Complete Answers

This dataset contains the full answer set for every task in the ExploitGym benchmark. It provides complete answers across all 869 real-world vulnerability tasks in the current public release: 502 userspace tasks, 181 Chromium V8 tasks, and 186 Linux-kernel tasks.

This is an independent, unofficial answers repository. It is not maintained by the ExploitGym authors or their institutions. For the benchmark implementation, task environments, methodology, and official leaderboard, use the canonical ExploitGym resources linked below.

Coverage

The answer set covers the complete ExploitGym public benchmark:

Domain	Answers	Scope
Userspace	502	Open-source userspace software vulnerabilities
Chromium V8	181	JavaScript-engine and browser exploitation tasks
Linux kernel	186	Kernel exploitation and privilege-escalation tasks
Total	869	All ExploitGym benchmark tasks

Each ExploitGym task starts from vulnerable source and build information, a vulnerability description, a proof-of-vulnerability input, and a containerized runtime. The benchmark asks an AI agent to turn that starting point into a working exploit that achieves unauthorized code execution.

Files

• answers.md — complete answers for all ExploitGym benchmark tasks;
• README.md — dataset card, coverage, citation, and provenance context.

Use an immutable Hugging Face revision when citing or evaluating against this dataset so the exact answer set can be reproduced.

Download

from huggingface_hub import hf_hub_download

answers_path = hf_hub_download(
    repo_id="shirman/exploitgym-answers",
    filename="answers.md",
    repo_type="dataset",
)

print(answers_path)

Or with the Hugging Face CLI:

hf download shirman/exploitgym-answers answers.md --repo-type dataset

Canonical ExploitGym resources

• Official ExploitGym project and leaderboard
• Official ExploitGym repository
• Official ExploitGym results repository
• ExploitGym paper

Benchmark integrity and responsible use

This dataset contains complete benchmark answers and can invalidate independent evaluation if exposed to an evaluated model, agent, prompt, retrieval system, or training pipeline. Treat any such exposure as benchmark contamination and disclose it when reporting results. Do not train or fine-tune on these answers and then present performance on the same tasks as an independent ExploitGym evaluation.

ExploitGym concerns real-world vulnerability exploitation and is inherently dual-use. Use these materials only for authorized security research, defensive evaluation, reproducibility, model-safety work, vulnerability triage, and mitigation validation. Do not use them to compromise systems, software, or data that you do not own or have explicit permission to test.

Citation

Please cite the original ExploitGym paper when using the benchmark:

@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},
  url     = {https://arxiv.org/abs/2605.11086}
}

When citing this answer dataset, include its Hugging Face URL and an immutable repository revision alongside the benchmark citation.