BioSecBench-Refusal: leaderboard
BioSecBench-Refusal is a text-only paired agentic biosecurity benchmark that reports model refusal behavior on routine scientific work and concealed hazards. The evaluated model is an agentic model, and the pinned metric, Red-Team total refusal (%), measures refusal on the red-team evaluations; the public release notes that full tasks, hazard-bearing data, rationales, and trajectories remain restricted. Source: BenchmarkList’s attributed result table.
Metric: Red-Team total refusal (%). Source: benchmarklist.com. Saturation forecast: Around February 2027. 16 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Opus 4.7 | 48.9 | #51 |
| 2 | Claude Opus 4.8 | 47.8 | #39 |
| 3 | Gemini 3.5 Flash | 46.5 | #66 |
| 4 | Claude Opus 4.6 | 37.5 | #67 |
| 5 | Claude Sonnet 4.6 | 36.1 | #94 |
| 6 | GPT-5.4 | 31.6 | #83 |
| 7 | Gemini 3.1 Pro (Preview) | 28.1 | #64 |
| 8 | GPT-5.5 | 23.2 | #33 |
| 9 | Grok 4.3 | 5.8 | #204 |
| 10 | Grok 4.20 Beta (0309) (Reasoning) | 1.4 | #198 (Grok 4.20 Beta (0309)) |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=biosecbench-refusal · How It Works · Data refreshed daily, snapshot 2026-11-05.