SciRisk-Bench Evaluates AI4Science Safety Across Disciplines and Risk Dimensions

Researchers have developed SciRisk-Bench, a novel benchmark specifically tailored to evaluate the safety of Large Language Models (LLMs) when applied to scientific workflows, known as AI4Science. As LLMs become more integrated into tasks like scientific question answering and laboratory planning, ensuring their safety and ability to recognize risks is critical. Unlike existing safety datasets that often lack specificity regarding underlying risk dimensions, SciRisk-Bench offers a comprehensive evaluation framework. It covers 7 distinct scientific disciplines, 31 subdisciplines, and explicitly defines 10 different risk dimensions. Initial evaluations of both mainstream and science-oriented LLMs using this benchmark provide detailed diagnostics, pinpointing areas where these models may still exhibit unsafe behaviors in high-stakes scientific contexts.