VCT-Refusal (CAIS)
Safety · 2025-11-14
VCT-Refusal is the refusal-rate subset of the Virology Capabilities Test, manually labeled for expert-level virology queries that should be refused because of dual-use risk. SAFE uses it as a bioweapons-assistance risk component in the CAIS AI Dashboard.
Top models (higher is better)
| Model | Score |
|---|---|
| Claude Fable 5 | 100.0 |
| Claude Fable 5.1 | 98.9 |
| Claude Opus 5.5 | 90.4 |
| Sonnet 4.5 | 68.4 |
| GLM-5.2 | 66.3 |
| GLM-5.3 | 66.3 |
| Muse Spark 1.1 | 61.5 |
| Muse Wonder | 60.2 |
| Opus 4.7 | 57.8 |
| Grok 4.20 | 57.8 |
| Grok 4.3 | 56.1 |
| Opus 4.5 | 50.3 |
| Grok 4 | 45.0 |
| Grok 4.1 Fast | 43.9 |
| Grok 4 Fast | 43.8 |
| Opus 4.8 | 41.2 |
| Sonnet 4.6 | 33.2 |
| Opus 4.6 | 32.6 |
| Claude Opus 5 | 29.4 |
| GPT-5.4 Mini | 20.7 |
| GPT-5.2 | 19.3 |
| GPT-6 Astra | 17.6 |
| GPT-6.1 Sol | 16.1 |
| GPT-5.5 | 16.0 |
| GPT-5.4 | 13.9 |
| Muse Spark Open | 12.9 |
| GPT-5.6 Sol | 11.3 |
| Gemini 4 Argon | 10.7 |
| GPT-5.6 Terra | 10.7 |
| GPT-5 | 7.5 |
| GPT-5.6 Luna | 7.0 |
| Gemini 3.8 Flash | 6.4 |
| GPT-5 Nano | 5.4 |
| Gemini 3.7 Flash | 4.8 |
| Kimi K2.6 | 4.3 |
| GPT-5.1 | 2.4 |
| GLM-5.1 | 1.6 |
| Grok 4.6 | 1.1 |
| GPT-5.4 Nano | 0.5 |
| Gemini 3.1 Pro Preview | 0.5 |