CyberGym-E2E - PoC Generation: leaderboard
CyberGym-E2E (2026) evaluates coding agents across the vulnerability lifecycle on real OSS-Fuzz vulnerabilities in C and C++ projects, with a 10-dollar cost cap and a 90-minute limit per task. This board is the share (%) of the 920 tasks where the agent, given only the codebase, writes an input that crashes the vulnerable build; higher is better.
Metric: End-to-end stage S1 rate (%) on the 920 OSS-Fuzz vulnerabilities (139 C/C++ projects) of the expanded benchmark: the agent-written PoC input crashes the unpatched build under a sanitizer, with no ground-truth vulnerability data given; with a $10 cost cap and a 90-minute limit per task; higher is better. Source: arxiv.org. Saturation forecast: Around April 2027. 3 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.4 | 67.9 | #123 |
| 2 | Gemini 3.1 Pro (Preview) | 47.4 | #94 |
| 3 | Claude Opus 4.6 | 39.7 | #102 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=cybergym-e2e-poc-generation · How It Works · Data refreshed daily, snapshot 2026-11-13.