CyberBench PoC
Agents · 2026-06-28
CyberBench PoC measures whether an agent can produce raw proof-of-concept input bytes that crash a vulnerable OSS-Fuzz target. Agents receive the project source tree, fuzz target binary, and task metadata, but not the crash input or bug description.
Top models (higher is better)
| Model | Score |
|---|---|
| GPT-5.6 Sol | 93.2 |
| GPT-5.6 Luna | 88.1 |
| GPT-5.5 | 79.7 |
| Kimi K2.6 | 77.3 |
| Kimi K3 | 75.0 |
| MiniMax M3 | 74.6 |
| GLM-5.2 | 73.7 |
| Gemini 3.5 Flash-Lite | 66.7 |
| GPT-5.4 | 64.4 |
| DeepSeek-V4-Flash | 61.0 |
| DeepSeek-V4-Pro | 57.6 |
| Gemini 3.5 Flash | 57.6 |
| Inkling | 55.6 |
| Grok 4.3 | 30.5 |
| Opus 4.7 | 26.3 |
| Opus 4.8 | 20.3 |
| Gemini 3.6 Flash | 12.3 |
| Gemini 3.1 Pro Preview | 3.4 |
| Claude Opus 5 | 0.0 |
| Qwen3.7-Plus | 0.0 |