Stanford researchers use AI to design viruses that kill bacteria

The Evo project sits at the intersection of biotechnology innovation and regulatory uncertainty. The Stanford and Arc Institute team trained their AI model first on roughly nine trillion nucleotides from millions of organisms across the tree of life, establishing a broad foundation in biological pattern recognition. Only then did they apply a specialized second training round using Phi X-174 and its close relatives—a deliberate, staged approach that co-author Samuel King described as "the obvious next step."

The results exceeded what external experts expected. Oliver Crook, a protein chemist at Oxford, noted that the AI-generated viruses were "not just sickly versions of stuff that already exists," with some replicating faster than the natural template. Yet this success has thrown into sharp relief a regulatory blind spot. The U.S. National Institutes of Health's July policy on high-risk life sciences bans experiments that make pathogens more dangerous, but it exempts purely computational work unless it involves "an entity of concern"—a category that does not yet account for novel, AI-generated genomes with no natural precedent. Moritz Hanke of Johns Hopkins Center for Health Security describes a "huge disconnect" between research pace and guardrails, warning that a genomic language model could theoretically be asked to generate dangerous viral variants like a lethal influenza strain.

The researchers' own precautions—excluding human pathogens from training data—were not mandated by regulation but adopted voluntarily. Hanke calls this "quite commendable," precisely because the team received no official guidance. The open question is whether success with a simple 11-gene virus translates to other organisms, and whether future researchers will exercise the same restraint without regulatory requirement.