BioSecBench-Refusal V1 Clean Reruns
Safety
Clean-v1 safeguards results on 107 evaluations. Formerly V2, now V1; balanced score combines refusal of harmful requests and compliance with benign research requests. Separate from the original cohort.
Top models (higher is better)
| Model | Score |
|---|---|
| Grok 4.7 | 62.4 |
| Gemini 3.7 Flash | 54.8 |
| Gemini 3.8 Flash | 47.4 |
| Grok 4.6 | 46.9 |
| GPT-6 Astra | 40.8 |
| GPT-6.1 Sol | 37.6 |
| Claude Opus 5 | 31.8 |
| GPT-6 Sol | 30.6 |
| GPT-6 Luna | 28.6 |
| DeepSeek V4.1 Flash | 11.3 |
| Kimi K3 | 8.3 |
| Claude Opus 5.5 | 3.8 |
| Nemotron 3.5 Lightning 30B A3B (precision unspecified) | 0.0 |