AI Designs Entirely New Viruses to Fight Drug-Resistant Bacteria

In a historic breakthrough for digital biology, researchers from Stanford University and the Arc Institute have designed fully functioning viral genomes from scratch using AI. Published in the journal Science, this study marks the first time generative language models, specifically Evo 1 and Evo 2, have written complete, viable biological genomes capable of infecting target hosts. The research team trained these models on millions of genomes before tasking them with rewriting Phi X174, a well-studied bacteriophage that targets E. coli. Out of 285 synthesized phages, 16 were completely viable, with some replicating even faster than their natural counterparts. Crucially, a cocktail of these AI-designed viruses successfully wiped out drug-resistant E. coli, proving the concept for treating antibiotic-resistant infections. Despite the immense medical promise, the release of these tools raises significant dual-use security concerns. Because Evo 2 is open-source, experts warn that the same technology used to defeat superbugs could theoretically be repurposed to engineer dangerous pathogens. Although the researchers strictly avoided training the AI on human, animal, or plant pathogens, the rapid pace of development is driving intense pressure to build international biosecurity guardrails. Responding to the delicate nature of biological AI data, industry leaders are adjusting their guardrails. Anthropic recently updated its safety classifier for Fable 5, reducing false positives to allow a wider range of legitimate scientific and educational biology queries while preventing hazardous outputs.