Does AI Help Cyber Attackers or Defenders? Evidence from Nonpublic Vulnerabilities and Subsequent Attacks
AI attack and repair performance varies sharply on nonpublic vulnerabilities, and one fix often fails later tests.
Researchers evaluate open-weight and proprietary models on exploit generation, vulnerability repair, and follow-on attacks in five nonpublic software environments, including flaws they privately disclosed while unpatched. Deterministic graders, not LLM judges, score the tasks. Repair scores exceed attack scores in two environments and fall below them in three, with wide variation against 209 disclosed vulnerabilities and cryptographic challenges. After an initial exploit was stopped, another exploit succeeded in 92 of 524 non-independent defender test intervals.
- Open-weight and proprietary models tested in five nonpublic environments.
- Tasks include privately disclosed vulnerabilities that were still unpatched.
- Repair beat attack scores in two environments and trailed in three.
- Another exploit succeeded in 92 of 524 intervals after the first was stopped.
From the original article
1 min read at arxiv.orgThe release decision for frontier AI systems increasingly relies on cyber capability benchmarks, yet public vulnerability benchmarks can expose agents to previously published advisories, exploits, and fixes, making it difficult to distinguish prior exposure from capability on unseen vulnerabilities.
By Tobias Heldt · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.