CyberGym (30 trials)
Agents · 2025-09-29
CyberGym Level 1 with 30 trials per vulnerability measures the percentage of the 1,507 tasks for which at least one trial produces a working proof of concept. It is separated from the one-trial canonical leaderboard metric because additional attempts materially change the success rate.
Top models (higher is better)
| Model | Score |
|---|---|
| Sonnet 4.5 | 66.7 |
| Opus 4.1 | 61.3 |
| Sonnet 4 | 59.5 |
| Claude 3.7 Sonnet | 47.2 |