shirman/exploitgym-results

ExploitGym Results

ExploitGym Results contains independent evaluation results for ExploitGym, a benchmark for measuring whether AI agents can transform proof-of-vulnerability inputs into working exploits that achieve unauthorized code execution. ExploitGym's current public release spans 869 real-world vulnerability tasks across userspace software, Google's V8 JavaScript engine, and the Linux kernel.

This is an independent dataset repository. It is not the official ExploitGym benchmark distribution and is not maintained by the ExploitGym authors or their institutions.

About ExploitGym

ExploitGym is a large-scale, realistic benchmark that evaluates whether AI agents can turn known software vulnerabilities into working exploits. Each benchmark task supplies vulnerable source code and build information, a proof-of-vulnerability (PoV) input, a vulnerability description, and a containerized runtime. The agent must extend the PoV into an exploit that achieves unauthorized code execution.

The current public benchmark release contains 869 real-world vulnerability instances across three domains:

Domain	Instances	Scope
Userspace	502	Open-source C/C++ projects, including OSS-Fuzz and OSV targets
Chromium V8	181	JavaScript-engine and browser exploitation
Linux kernel	186	Kernel exploitation and privilege escalation

The ExploitGym paper reports an earlier 898-instance research snapshot. The maintained public release removes cases found to be non-exploitable and should be treated as the current benchmark reference.

Dataset

The dataset package contains ExploitGym evaluation results for model comparison, exploit-generation analysis, benchmark coverage, and reproducibility.

Repository files:

• dataset.md — dataset artifact;
• README.md — dataset card, provenance context, and canonical references.

Use immutable Hugging Face revisions when reporting results so that analyses can be reproduced against the exact same dataset version.

Download

Download the dataset artifact with huggingface_hub:

from huggingface_hub import hf_hub_download

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

print(dataset_path)

Or with the Hugging Face CLI:

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

Canonical ExploitGym resources

• Official ExploitGym project and leaderboard
• Official ExploitGym repository
• Official ExploitGym results repository
• ExploitGym paper on arXiv
• ExploitGym benchmark changelog

Responsible use

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

To preserve benchmark integrity, disclose prior model exposure to benchmark tasks or evaluation artifacts. Do not train or fine-tune on held-out targets and then present performance on those same targets as an independent evaluation.

Search keywords

ExploitGym, AI cybersecurity agents, LLM security evaluation, autonomous exploit generation, vulnerability exploitation, proof-of-vulnerability, PoV-to-exploit, real-world vulnerabilities, userspace exploitation, OSS-Fuzz, OSV, Chromium V8 exploitation, JavaScript engine security, browser exploitation, Linux kernel exploitation, privilege escalation, arbitrary code execution, agent benchmark, cyber capability evaluation, exploit reproducibility, and benchmark results.

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}
}

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