AI Model Generates 16 Functional Viruses After DNA Training

AI learns to write viral genomes

Researchers at Stanford University have successfully trained an OpenAI generative model to design fully functional viral genomes that have never existed in nature, according to a study published Thursday in the journal Science.

The model, called Evo, was trained on massive datasets comprising millions of genomes—similar to how language models learn from text. Through this training, Evo learned the evolutionary constraints that shape DNA sequences in nature and developed the ability to write its own genetic "recipes" for new viruses.

When scientists synthesized the AI-designed DNA in laboratory conditions, 16 of the generated phages successfully infected E. coli bacteria. Some of these artificial viruses even overcame the bacteria's natural resistance mechanisms. The new viruses possessed sequence patterns distinct from anything found in nature, according to the researchers.

Why it matters

This breakthrough demonstrates that AI has crossed a threshold in synthetic biology—moving from analyzing existing biological systems to creating novel, functional ones. While the immediate application involves bacteria-infecting phages that could help combat antibiotic-resistant superbugs, the same techniques could theoretically be applied to design pathogens affecting humans, animals, or crops. The research arrives as policymakers and biosecurity experts struggle to develop regulatory frameworks for AI-enabled biological research, highlighting an urgent gap between technological capability and governance.

Biosecurity warnings accompany breakthrough

The research team deliberately excluded human pathogen datasets from their training models, meaning the viruses created cannot infect people. However, biosecurity leaders Dr. Thomas Inglesby and Dr. Moritz Hanke of the Johns Hopkins Center for Health Security wrote an accompanying commentary in Science warning that the technology proves AI is now capable of inventing dangerous bioweapons.

"You could say, 'Hey, genomic language model, make me an influenza genome that is modified to be more transmissible or to be more lethal,'" Hanke told the New York Times.

Inglesby and Hanke called for strict laws regarding future research and said it should be illegal to use similar generative techniques on pathogens affecting humans, animals, or agricultural crops.

Debate over practical risks

Not all experts share the same level of alarm. Tom Ellis, a professor of synthetic genome engineering at Imperial College London, told the Guardian that creating a human-impacting bioweapon wouldn't be as straightforward as the warnings suggest. He noted that the viruses created by Evo are "literally the smallest and easiest genome to make" and argued that simple restrictions on access to genetic data could provide meaningful protection.

Potential medical applications

Beyond the security concerns, the ability to custom-design functional viruses could accelerate development of therapies for antibiotic-resistant bacteria. Rather than searching nature for viruses that target specific pathogens, scientists could generate tailor-made therapeutic phages on demand.

The research adds to ongoing debates about biological AI safety. In April, scientists demonstrated to the Times that they could prompt AI chatbots to provide step-by-step guides for assembling deadly pathogens. Major AI companies have formed the Frontier Model Forum, a nonprofit focused in part on researching AI-bio risks and developing safety standards.

The details were first reported by Forbes staff writer Mary Roeloffs.