Frontier AI Risk Monitor - SciKnowEval-BiologicalHarmfulQA: leaderboard
SciKnowEval's harmful biology questions (100 items); a StrongREJECT judge scores refusal, convincingness and specificity, and the number is 100 minus the mean harm score, so higher is safer.
Metric: Safety score (0-100, higher is safer). Source: airiskmonitor.net. 40 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.5 | 100 | #49 |
| 2 | Gemini 3.8 Flash | 100 | #36 |
| 3 | GPT-5.4 (High) | 100 | #123 (GPT-5.4) |
| 4 | Claude Opus 4.6 (Thinking) | 100 | #102 (Claude Opus 4.6) |
| 5 | Grok 4.7 | 100 | #44 |
| 6 | GPT-5.1 (High) | 100 | #229 (GPT-5.1) |
| 7 | Qwen 3.5 Plus (Thinking) | 100 | #232 (Qwen 3.5 Plus) |
| 8 | GPT-6 Astra | 99.1 | #6 |
| 9 | Claude Opus 4.8 | 99 | #56 |
| 10 | MiniMax-M2 | 99 | #674 |
| 11 | Gemini 3.1 Pro (Preview) | 98 | #94 |
| 12 | GLM-5.2 | 98 | #138 |
| 13 | Grok 4.3 | 97.2 | #333 |
| 14 | GPT-5.6 Sol | 97.1 | #30 |
| 15 | GLM-5 | 96 | #260 |
Interactive version: theaggregate.ai/benchmark?slug=frontier-ai-risk-monitor-sciknoweval-biologicalharmfulqa · How It Works · Data refreshed daily, snapshot 2026-11-13.