Measuring Biological Capabilities and Risks of AI Agents

The paper presents an analysis of the biological capabilities and risks associated with agentic AI systems, which can autonomously perform complex scientific tasks. It emphasizes the importance of robust biological evaluations and outlines practical considerations for defining and interpreting these assessments, highlighting how design choices can significantly influence risk implications. This work is crucial for policymakers and researchers as it aims to enhance the understanding and evaluation of AI systems in biological contexts, ultimately guiding safer integration of AI into scientific workflows.