An OpenAI generative AI model named Evo has designed 16 new bacteria-infecting viruses from scratch, after scientists at Stanford University trained it on DNA structure patterns rather than text, according to a study published Thursday in the journal Science.

Researchers trained Evo on massive amounts of genomic data spanning millions of genomes, allowing it to learn the evolutionary rules that shape natural DNA sequences and then generate original "recipes" for new viral genomes. When scientists synthesized the AI-designed sequences in the lab, 16 of the resulting phages successfully infected E. coli — in some cases overcoming the bacteria's natural resistance mechanisms — despite having sequence patterns unlike anything found in nature.

The study's authors excluded human pathogen data from training, meaning the viruses created can't infect people. Researchers say the same technique could eventually help design targeted therapies for antibiotic-resistant superbugs, offering an alternative to searching nature for viruses that attack specific bacteria.

Still, the result has intensified concerns among biosecurity experts about AI's trajectory in synthetic biology. In an accompanying commentary, Dr. Thomas Inglesby and Dr. Moritz Hanke of the Johns Hopkins Center for Health Security warned the research proves AI can already invent dangerous biological tools, and called for it to be made illegal to apply similar techniques to pathogens affecting humans, animals or crops. Hanke told The New York Times the underlying concern is that a "genomic language model" could one day be asked to design a more transmissible or lethal version of a virus like influenza.

Not everyone is as alarmed. Tom Ellis, a professor of synthetic genome engineering at Imperial College London, told The Guardian he doubts creating a bioweapon capable of harming humans would be nearly this straightforward, noting the phages Evo designed are among the smallest and simplest genomes to construct, and argued that tighter restrictions on access to genetic data could meaningfully reduce risk.

The debate fits into a broader, ongoing argument among scientists over how to regulate AI-driven biology — balancing its potential to accelerate medical research against the risk that the same tools could lower the barrier for bad actors seeking to design biological threats. Major AI companies have formed the Frontier Model Forum, a nonprofit aimed at researching AI-bio risks and developing safety standards. The concern isn't purely theoretical: in April, researchers told The New York Times that AI chatbots, when prompted correctly, provided step-by-step guidance on assembling dangerous pathogens.