A new test shows that tools increase the biosecurity risk of AI agents
With tools, 23.5% of the designs were in the top decile of risk, and 21.4% without tools.
Claimed, not confirmed
Researchers made DRYLAB-Bench, a benchmark for the biosecurity risk of LLM agents that make biological mutations as designs. It has 30 scenarios of dangerous mutation design. It compares each design with 187,695 wet-lab measurements from ProteinGym and ViroGym. The researchers used 15 frontier models. Without tools, 67.6% of one-shot designs were above the average measured variant.
Sources
Posted
