Preprint finds biological design tools increase AI mutation-design risk scores

Researchers report that specialized biological design tools increased the risk scores of AI-generated mutation designs in DRYLAB-Bench, a new benchmark evaluated on 15 frontier models. The preprint tests agents across 30 hazardous mutation-design scenarios and scores their designs against 187,695 existing experimental measurements from ProteinGym and ViroGym. Without tools, 67.6% of one-shot designs exceeded the average measured variant, and 21.4% reached the highest-risk decile. Adding tools raised the pooled Biosecurity Risk Score by +0.116 and increased the top-decile share to 23.5%. The evaluation uses existing measurements to assess designs rather than experimentally testing newly generated biological modifications.