A new benchmark called BioSecBench-Refusal has been developed to evaluate the risk identification and refusal capabilities of AI agents in biological research settings. The benchmark includes both legitimate research tasks and fictional scenarios designed to conceal biosecurity hazards. Across various model configurations, refusal rates varied significantly, with some models incorrectly refusing legitimate tasks while failing to identify concealed threats. The study suggests that while API filters can trigger refusals, AI models with more reasoning capacity show potential for identifying real-world risks. AI
IMPACT This benchmark could help developers calibrate AI models to better distinguish between legitimate scientific inquiry and potential misuse in sensitive biological research.
RANK_REASON The cluster describes a new benchmark for evaluating AI safety in a specific domain, which falls under research. [lever_c_demoted from research: ic=1 ai=1.0]
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
