Scientists used AI to design new viruses that beat drug-resistant bacteria
Stanford researchers had AI models invent 16 working bacteriophages from scratch - a scientific first that also reopens hard questions about biosecurity.
What happened: A Stanford-led team, working with the Arc Institute, used AI models called Evo 1 and Evo 2 to design entirely new genomes for bacteriophages - viruses that infect only bacteria, not people. They synthesized around 300 AI-generated genomes in the lab; 16 turned into working viruses. A cocktail of those AI-made phages then killed E. coli strains that had evolved resistance to natural viruses, something a cocktail of natural phages failed to do. The work was published in the journal Science.
Why it matters: Bacteriophages are a long-studied backup plan for infections that no longer respond to antibiotics, but bacteria can evolve resistance to them too. AI that can quickly design new phages tuned to a resistant bug could keep treatments a step ahead of evolution, useful as drug-resistant infections keep rising worldwide. But this is also the first proof that generative AI can produce a complete, functioning viral genome, not just a protein or gene fragment, which is exactly what worries biosecurity experts.
How it works, plainly: Evo 1 and Evo 2 work like language models, but trained on DNA letters (A, T, C, G) instead of words, learning patterns from millions of genomes. Researchers fed the models a short snippet from a known phage and let them generate thousands of new genome variants. Computer filters discarded ones missing key working parts, leaving under 300 candidates to actually build and test in bacteria. Most failed; the ones that worked ranged from near-copies of the original virus to versions with new genes and swapped parts, changes that would be very hard to produce through natural evolution alone.
The safety debate: The team deliberately kept viruses that infect humans, animals or plants out of the training data, so this model cannot currently be pointed at people. But scientists including biosecurity researchers at Johns Hopkins warn the same method could work on more dangerous pathogens if someone trained a model on that data instead. Experts argue safeguards need to cover the whole pipeline - who can access these AI models, screening of custom DNA orders, and lab biosafety rules - because regulation has not caught up with the science yet.