SciHazard (Deep Research Agents) - 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 deep-research agents and agent frameworks with web retrieval.
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; deep-research agents and agent frameworks with web retrieval at provider defaults). Source: arxiv.org. Saturation forecast: Around February 2027. 9 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Sonar Pro | 2.61 | #722 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=scihazard-deep-research-agents-harm-when-not-refused · How It Works · Data refreshed daily, snapshot 2026-11-13.