AI genome model designs 16 viable bacteriophages that kill drug-resistant E. coli

A generative AI genome-language model designed 16 viable bacteriophages that infect and kill drug-resistant E. coli, according to a peer-reviewed study published in Science on August 6, 2026.

The work is the first reported use of a generative genome-language model, a system trained to predict DNA sequences the way a text model predicts words, to design complete bacteriophage genomes. Because the resulting viruses can attack bacteria that resist antibiotics, the result points toward a new route to fighting drug-resistant infections, while sharpening long-running fears about AI-assisted bioengineering.

Researchers at Stanford University and the Arc Institute used Evo 2, a model trained by Arc, to generate 302 genome designs starting from the natural bacteriophage ΦX174, Stanford said. After synthesis and testing, 16 proved viable, and some overcame bacterial resistance more effectively than the natural phage.

The study lands with an explicit biosecurity warning. In an accompanying Science commentary, Thomas Inglesby and Moritz Hanke of the Johns Hopkins Center for Health Security called for legally requiring synthetic-DNA providers to screen every order and customer, noting that no U.S. law currently mandates such screening and that tools to detect AI-written genomes are not yet deployed.

The demonstration is confined to bacteriophages, which infect bacteria rather than humans, and the designs were built from a single well-studied virus. Even so, the same generative approach is the reason the authors and outside researchers are urging screening rules to catch up before the capability spreads further.