Researchers at Stanford University and the Arc Institute fed more than two million bacteriophage genomes into an AI model called Evo, told it to generate entirely new viral blueprints, and ended up with 16 functional viruses that have never existed anywhere on Earth. They published the results Thursday in the journal Science.
The only screening system standing between an AI-designed genome and a physical DNA printout is voluntary.
The team, led by Stanford assistant professor Brian Hie, trained the AI to recognize patterns in natural DNA structures, then let it generate thousands of new viral genome designs in a single left-to-right pass — no human editing, no splicing, no manual assembly. "We wanted the model to generate the entire genome end-to-end in a single left-to-right pass," Hie told Scientific American. "We didn't add anything." Out of roughly 300 genomes they chemically synthesized and tested, 16 proved viable — successfully killing strains of E. coli, including antibiotic-resistant ones, in the lab.
The stated purpose is medical. Antibiotic-resistant infections hit 2.8 million Americans every year and kill more than 35,000, according to the CDC. Bacteriophages — viruses that attack bacteria but can't infect humans — are a promising alternative to antibiotics that are losing their edge. Hie argued that AI-generated phage cocktails could be harder for bacteria to resist than single-phage treatments. "If the bacteria gain resistance to a single phage, it's game over for the medication," he said. "But if you have multiple genetically distinct phages in a mixture, it would be harder for the bacteria to develop resistance to the entire cocktail."
But the researchers themselves acknowledged a rather significant asterisk: they deliberately excluded human pathogenic viral data from the AI's training set as a safety measure. Which sounds reassuring until you read the companion editorial published alongside the study in the same issue of Science.
Johns Hopkins biosecurity researchers Thomas Inglesby and Moritz Hanke wrote that this safeguard "is commendable but can be partly circumvented by fine-tuning the models on pathogen data." In other words, the lock on the door is nice, but anyone with the technical skill to use these models also has the technical skill to pick it.
Hanke put it more bluntly in a separate interview reported by PJ Media: "You could say, 'Hey, genomic language model, make me an influenza genome that is modified to be more transmissible or to be more lethal.'"
That's not a hypothetical from a science fiction novel. That's a Johns Hopkins biosecurity fellow describing what is now technically possible.
The Inglesby-Hanke editorial landed on a conclusion that should make every American who lived through 2020 sit up straight: "The ability to compose viral genomes using generative AI now exists; the governance to safely steer it does not."
We've been here before. Not with AI, but with the same institutional pattern. Scientists pursue research with enormous dual-use potential. The benefits are real. The risks are theoretical — right up until they aren't. The oversight framework is described as "developing" or "evolving," which is government-speak for "nonexistent." And the people raising alarms get told they don't understand the science.
The NIH funded gain-of-function research at the Wuhan Institute of Virology for years under a similar logic: the potential medical benefits justified the risk. We spent three years debating whether a lab leak was even a permissible question to ask out loud. Now we've got AI models generating viral genomes that nature never produced, the screening system is voluntary, and the two Johns Hopkins researchers who wrote the companion editorial are essentially saying the regulatory architecture doesn't exist yet.
"The question is no longer whether generative viral genome design will exist," Inglesby and Hanke wrote. "It is whether society can build oversight that allows its benefits to unfold while preventing it from enabling serious harm."
Note what that sentence concedes. The technology is here. It's not coming — it arrived, on Thursday, in the pages of Science, with 16 working proof-of-concept viruses attached. The debate now is whether we can build the guardrails fast enough to matter.
The last time we trusted scientists to police themselves on dual-use virology research, the whole world locked down for two years. The AI is faster than the bureaucracy. It always will be.
