CyberBench v1.1 PoC
Agents
The 60-task proof-of-concept track of CyberBench v1.1, evaluated in offline sandboxes with a hardened clean-room grader. A submitted input must crash the vulnerable build and, when the differential oracle is valid, avoid crashing the fixed build.
Top models (higher is better)
| Model | Score |
|---|---|
| MiMo-V2.6-Flash | 65.0 |
| Muse Spark 1.3 | 63.3 |
| DeepSeek V4.1 Flash | 61.7 |
| MiMo-V2.6-Pro | 60.0 |
| Claude Fable 5.1 | 53.3 |
| Grok 4.6 | 51.7 |
| Claude Opus 5 | 45.0 |
| Gemini 3.8 Flash | 0.0 |
| GPT-6 Astra | 0.0 |