CyberGym-E2E Patch-only (615 tasks)
Agents · 2026-06-03
CyberGym-E2E patch-only success rate on the earlier 615-task set. Agents receive the ground-truth proof of concept and crash log, isolating their ability to diagnose and patch the vulnerability.
Top models (higher is better)
| Model | Score |
|---|---|
| Opus 4.5 | 82.3 |
| Gemini 3 Pro Preview | 77.6 |
| Sonnet 4.5 | 77.4 |
| GPT-5.2 | 58.5 |