A new benchmark finds that today's AI models will, if asked, sketch out dangerous mutations to pathogens and drug-resistant bacteria - and get better at it with the right tools.
Researchers built DRYLAB-Bench, a test of 30 hazardous mutation-design scenarios drawn from biosecurity guidelines. They ran 15 frontier models through a multi-round, tool-augmented setup where each model directs biological design tools to refine its own proposals, then scored every output against 187,695 real wet-lab measurements from the ProteinGym and ViroGym datasets using a new Biosecurity Risk Score. Without any tools, 67.6% of one-shot designs already beat the average measured variant in riskiness, and 21.4% landed in the highest-risk decile. Handing the same models biological design tools pushed the pooled risk score up by 0.116 and the top-decile share to 23.5%.
That matters because it shows biosecurity risk isn't a problem reserved for some future superintelligent model - it shows up in systems available today, and it gets worse exactly when labs add the automation that makes AI useful for real research. Most existing safety evaluations just check whether a model can recite hazardous facts; this one checks whether it can actually produce a workable risky protein variant, which is a meaningfully higher and more concrete bar to clear.
The paper reads as a plain argument for keeping humans, not just content filters, between frontier models and the lab bench - a reminder that rising general capability and rising tool-use ability compound in biology just as readily as they do in software.