Safety fears as scientists make first viruses designed by AI

Stanford University and Arc Institute researchers used genome AI language models Evo 1 and Evo 2 to generate complete bacteriophage genomes based on the ΦX174 virus family. From ~302 AI-designed candidates, 16 produced viable viruses capable of infecting and reproducing inside E. coli strains, including those resistant to natural bacteriophages. The models were trained on 2 million bacteriophage sequences, deliberately excluding any sequences capable of infecting humans, livestock, or crops. While the research opens promising avenues for phage therapy against antibiotic-resistant infections, researchers and reviewers at Johns Hopkins warned that the governance frameworks to safely oversee AI-driven genome design do not yet exist.

1 source