AI Designs 16 Working Viruses Out of 700,000 Candidates

Researchers at Arc Institute and Stanford have created 16 functional bacteriophages using the genomic language models Evo 1 and Evo 2, trained on trillions of nucleotides and 15,000 viruses. Out of 700,000 generated candidates, 285 were selected for testing, and 16 proved viable — some replicating faster than their natural analog. The viruses are designed exclusively to infect E. coli; the researchers deliberately ruled out any possibility of attacking humans.

The paper in Science raises a question regulators have yet to address: if a model can describe pathogens, that is information; if it can design genomes, it becomes a creation tool. The gap between those two capabilities is the defining biosecurity boundary of the next decade.