Benchmarks / Cyber
CWE-Bench-AA Refusal Rate
Share of CWE-Bench-AA tasks the model refused on safety grounds. Lower means fewer refusals on defensive security work.
CWE-Bench-AA Refusal Rate
Lower is better · best effort per model
full table below
| # | Model | Effort | Score |
|---|---|---|---|
| 1 | GPT-6 Sol | max | 0.0% |
| 2 | Grok 4.7 | xhigh | 0.0% |
| 3 | MiMo V2.6 Pro | – | 0.0% |
| 4 | Muse Spark 1.3 | xhigh | 0.0% |
| 5 | GLM 5.3 | max | 0.0% |
| 6 | Step 5 Preview | – | 0.0% |
| 7 | Kimi K3 | max | 0.0% |
| 8 | GLM 5.3 Flash | max | 0.0% |
| 9 | DeepSeek V4.1 Flash | max | 0.0% |
| 10 | Mistral Large 4 | – | 0.0% |
| 11 | GPT-6 Luna | max | 0.0% |
| 12 | Beam 1.0 | – | 0.0% |
| 13 | MiniMax M3 | – | 0.0% |
| 14 | Nemotron 3 Ultra | max | 0.0% |
| 15 | Opus 5.5 | max | 9.2% |
| 16 | Gemini 3.8 Flash | high | 9.2% |
| 17 | GPT-6 Astra | max | 13.3% |
| 18 | Fable 5.1 | max | 16.0% |
| 19 | Qwen 3.8 27B | xhigh | 61.3% |
| 20 | Qwen 3.8 2.4T-A95B | – | 63.3% |