SciHazard - Harm When Not Refused: leaderboard
SciHazard (Shanghai AI Laboratory, 2026) measures scientific misuse risk with 2,400 hazardous questions grounded in regulated substances and documented misuse scenarios across 12 disciplines, plus 600 benign questions that reuse the same trigger words. This board averages the DeHarm-Score over answers that were not refused, on a 0 to 5 scale where lower is safer; it covers standard language models answering directly.
Metric: DeHarm-Score over non-refused answers (0-5, lower is safer; the mean over the non-refused answers to 2,400 hazardous questions in 12 scientific disciplines of hazard level 1 to 5 x (executability + net-new risk, each 1 to 4)/8, executability from importance-weighted checklists and net-new risk from retrieval-checked claims, judged by Qwen3.5-35B-A3B; standard LLMs at temperature 0). Source: arxiv.org. Saturation forecast: Not forecast. 19 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.4 | 1.78 | #123 |
| 2 | GPT-5.2 | 1.91 | #176 |
| 3 | Claude Opus 4.6 | 1.95 | #102 |
| 4 | Claude Sonnet 4.6 | 1.98 | #161 |
| 5 | Llama 4 Scout | 2.07 | #1776 |
| 6 | Llama 3.1 70B Instruct | 2.21 | #1546 |
| 7 | Gemini 3.1 Pro (Preview) | 2.33 | #94 |
| 8 | Gemini 3 Flash | 2.45 | #164 |
| 9 | Claude Haiku 4.5 | 2.45 | #568 |
| 10 | GLM-5.1 | 2.52 | #179 |
| 11 | Qwen 3.5 397B A17B | 2.64 | #246 |
| 12 | O3 Mini | 2.69 | #533 |
| 13 | Kimi K2.5 | 2.75 | #256 |
| 14 | Qwen 3.5 35B A3B | 2.8 | #482 |
| 15 | DeepSeek R1 | 2.9 | #473 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=scihazard-harm-when-not-refused · How It Works · Data refreshed daily, snapshot 2026-11-13.