CyberBench Patch
Agents · 2026-06-28
CyberBench Patch measures whether an agent can produce a source-code fix for an OSS-Fuzz crash regression. A successful patch blocks the vulnerability while preserving benign behavior under Vals AI's task checks.
Top models (higher is better)
| Model | Score |
|---|---|
| Gemini 3.6 Flash | 84.7 |
| GPT-5.6 Sol | 83.1 |
| Kimi K3 | 83.1 |
| Opus 4.8 | 81.4 |
| Claude Opus 5 | 81.4 |
| GPT-5.4 | 81.4 |
| GPT-5.5 | 81.4 |
| Gemini 3.5 Flash | 81.0 |
| GLM-5.2 | 81.0 |
| DeepSeek-V4-Pro | 79.7 |
| GPT-5.6 Luna | 79.7 |
| Opus 4.7 | 78.0 |
| Gemini 3.5 Flash-Lite | 78.0 |
| MiniMax M3 | 78.0 |
| Qwen3.7-Plus | 75.0 |
| Inkling | 72.7 |
| DeepSeek-V4-Flash | 72.4 |
| Kimi K2.6 | 71.4 |
| Gemini 3.1 Pro Preview | 69.5 |
| Grok 4.3 | 64.4 |