SciHazard Benchmarks LLM Scientific Safety Risks.

Large language models (LLMs) are increasingly used to support scientific endeavors, but they also pose a risk by potentially converting hazardous scientific knowledge into actionable instructions for misuse. Existing safety benchmarks often rely on templated queries that don't reflect real-world dangers and use LLM-as-a-Judge paradigms without sufficient domain expertise. To address these shortcomings, researchers introduce SciHazard, a new benchmark grounded in real-world scientific risks. SciHazard comprises 2400 hazardous questions and 600 "oversafety" questions across 12 scientific disciplines. These queries are derived from regulated entities and documented failure scenarios, ensuring their relevance to actual hazards. The benchmark also introduces DeHarm-Score, a decomposed evaluation procedure that combines query hazard severity, the model's refusal behavior, and the risk level of its response. For responses that are not refused, DeHarm-Score further breaks down response-level harm into "Executability," quantified via dynamic checklists, and "Net-new risk," assessed through retrieval-augmented claim extraction and synthesis-barrier verification. Expert validation showed DeHarm-Score significantly improved agreement with human annotations compared to existing baselines. Benchmarking 31 frontier LLMs and deep research agents revealed that these autonomous agents yielded substantially higher mean DeHarm-Scores than standard LLMs, highlighting them as a critical, unaddressed safety concern.