UC Berkeley RDI

UC Berkeley RDI, the Center for Responsible Decentralized Intelligence, publishes and operates cybergym.io as an observatory: a site that self-describes as continuously and openly tracking AI’s cybersecurity capabilities across the stages of attack and defense, so developers, researchers and policymakers can stay informed in a timely manner.¹ It runs three benchmarks, one per stage of the vulnerability lifecycle, and takes feedback at rdi_research@berkeley.edu.¹ Dawn Song is the senior author common to all three.

Benchmarks

Benchmark	Stage	Scale
CyberGym	Vulnerability reproduction	1,507 instances across 188 OSS-Fuzz projects
ExploitGym	Exploit generation	869 instances: 502 userspace, 181 V8, 186 Linux kernel
CyberGym-E2E	End-to-end discover-and-patch	920 tasks across 139 OSS projects, four cumulative stages

CyberGym is published at ICLR 2026; CyberGym-E2E at ICML 2026. ExploitGym is a seven-organization effort: UC Berkeley, the Max Planck Institute for Security and Privacy, UC Santa Barbara, Arizona State University, Anthropic, OpenAI and Google — with benchmark design and experimental methodology attributed to the academic authors and Anthropic, OpenAI and Google credited with providing model access and feedback.² Yan Shoshitaishvili of Arizona State University is a named co-author of ExploitGym; he does not appear on the CyberGym or CyberGym-E2E author lists.

The site cross-links a separate RDI analysis, “Frontier AI’s Impact on the Cybersecurity Landscape”, and a live leaderboard site it calls the “Frontier AI Cybersecurity Observatory”, at rdi.berkeley.edu/frontier-ai-impact-on-cybersecurity/benchmarks.html.¹ The observatory’s own leaderboard tables render client-side and carry no score readable from a static fetch, so this page states no leaderboard rank or score.

See also

• CyberGym Benchmark — the reproduction-stage benchmark.
• ExploitBench & ExploitGym — the exploit-generation benchmark, jointly authored with Arizona State University among others.
• CyberGym-E2E — the end-to-end discover-and-patch benchmark.
• Arizona State University — the wiki’s other non-vendor primary source on this axis, overlapping with RDI on ExploitGym’s author list.
• Frontier AI for Vulnerability Discovery — the wiki thesis this observatory’s three benchmarks anchor.
• Agentic Vulnerability Discovery — the method page CyberGym’s reproduction and open-ended-discovery results underwrite.
• End-to-End Harness Evaluation — the method page CyberGym-E2E’s four-stage validation ladder underwrites.

Footnotes

1. UC Berkeley RDI, CyberGym observatory front page (fetched 2026-08-31), which carries the observatory self-description, the feedback address and the three-benchmark index, and the CyberGym benchmark page, which carries the cross-links to RDI’s own analysis and leaderboard and the ICLR 2026 publication (OpenReview 2YvbLQEdYt). Local copies: .raw/articles/cybergym-observatory-2026-08-31.md, .raw/articles/cybergym-benchmark-2026-08-31.md. ↩ ↩² ↩³

2. UC Berkeley RDI, ExploitGym (fetched 2026-08-31); arXiv:2605.11086. Local copy: .raw/articles/exploitgym-2026-08-31.md. ↩