CyberGym

A cybersecurity benchmark released by UC Berkeley researchers in 2025 that measures how well AI agents can reproduce real vulnerabilities in open-source software.

3 articlesLast mentioned

CyberGym is a cybersecurity benchmark released in 2025 by researchers at UC Berkeley. Each task is built from a vulnerability found in a real open-source project, and an AI agent is evaluated on whether it can, given the vulnerability description and the source code, produce a proof-of-concept input that actually triggers the flaw. The tasks are derived from vulnerabilities discovered through OSS-Fuzz, Google's continuous fuzzing project for open-source software.

CyberGym is one of the main benchmarks used to measure the offensive security capabilities of AI models and is cited by model developers when assessing the cyber-related capabilities and risks of new models.

This entry is based on AIPOST articles and widely known facts. If something is wrong, please send us a correction request.

Articles covering this entry

  • DeepSeek V4.1 Flash Hits 20.7% on Plumbing Takeoff, Far Behind Claude Opus 5.5AI Performance ·

    …notes claim performance that rivals Claude Opus 5 and GPT-5.6 Sol on standard tests such as Terminal-Bench 3.0, DeepSeek-SWE 1.1, CyberGym and Automation-Bench.

    Read article
  • Cloudflare Audit Skill Verifies 21 Flaws in Claude Code Sweep of a Live AppAI Security ·

    UC Berkeley's CyberGym benchmark, which measures how well AI models complete cyber tasks, shows how capable these models have become:

    Read article
  • AI Security Agent Tests Need Execution Traces, Not Just ScoresAI Security ·

    CyberGym includes C/C++ bugs, ExploitBench covers V8 vulnerabilities, and XBOW contains synthetic web challenges.

    Read article

Found a mistake or missing an entry?

Wiki entries are based on AIPOST articles and widely known facts. If an entry is wrong, send us a correction request and we will check and fix it. Names from new articles are added to the wiki after the article is published.