LLM agents now exceed human expert performance on practical biology tasks, including dual-use capabilities like DNA synthesis evasion, raising urgent questions about how to safely deploy AI in biological research.
This paper introduces ABC-Bench, a benchmark that measures how well AI agents can perform biology tasks—both helpful ones like DNA design and risky ones like evading safety screening. Researchers tested multiple LLM agents and found they outperformed expert humans on all tasks, with some agents successfully writing working code for lab robots.