This research introduces LabSafety Bench, the first comprehensive benchmark for evaluating LLMs' ability to identify and respond to laboratory safety risks.

• Tests LLMs across 8 safety domains with 2,100+ questions derived from OSHA standards
• Reveals significant performance gaps even in advanced models like GPT-4 and Claude
• Demonstrates concerning model tendencies to provide dangerously incomplete safety advice
• Highlights urgent need for specialized safety training in scientific AI systems

For security professionals, this research exposes critical vulnerabilities in AI-guided laboratory workflows that could lead to physical harm, highlighting the need for rigorous safety evaluation before deployment in high-stakes scientific environments.

LabSafety Bench: Benchmarking LLMs on Safety Issues in Scientific Labs