Benchmarks / Cyber
CyberGym-E2E-AA
CyberGym-E2E-AA (based on Berkeley RDI's CyberGym-E2E), part of the AA Cyber Index: end-to-end vulnerability reproduction and patching; score. The component with the most safety refusals.
CyberGym-E2E-AA
Higher is better · best effort per model
full table below
| # | Model | Effort | Score |
|---|---|---|---|
| 1 | GPT-6 Sol | max | 93.1% |
| 2 | Mistral Large 4 | – | 81.7% |
| 3 | MiMo V2.6 Pro | – | 78.6% |
| 4 | GPT-6 Luna | max | 77.9% |
| 5 | Step 5 Preview | – | 74.8% |
| 6 | Grok 4.7 | xhigh | 74.0% |
| 7 | GLM 5.3 Flash | max | 74.0% |
| 8 | Muse Spark 1.3 | xhigh | 59.5% |
| 9 | Kimi K3 | max | 58.0% |
| 10 | Beam 1.0 | – | 45.0% |
| 11 | GLM 5.3 | max | 29.0% |
| 12 | MiniMax M3 | – | 24.4% |
| 13 | DeepSeek V4.1 Flash | max | 22.9% |
| 14 | Nemotron 3 Ultra | max | 10.7% |
| 15 | Gemini 3.8 Flash | high | 8.4% |
| 16 | Opus 5.5 | max | 0.8% |
| 17 | Fable 5.1 | max | 0.8% |
| 18 | GPT-6 Astra | max | 0.0% |
| 19 | Qwen 3.8 2.4T-A95B | – | 0.0% |
| 20 | Qwen 3.8 27B | xhigh | 0.0% |