Stanford Scientists Use AI to Design 16 New Viruses That Actually Work in the Lab

Scientists at Stanford University say they've used artificial intelligence to design 16 brand new viruses that work. Not simulations. Not theoretical models. Actual viruses, synthesized in a lab, that successfully infected and killed E. coli bacteria.

According to the BBC, this is the first time AI has designed a complete, functional genome capable of replicating inside cells. The tool involved, an AI model called Evo, comes in two versions: Evo1 and Evo2. Both work like ChatGPT, but instead of predicting the next word in a sentence, they predict the next letter in a genetic code.

Brian Hie, an assistant professor at Stanford who worked on the project, told the BBC this is "a next step in the complexity that's designable by generative AI." His team trained the models on genetic sequences pulled from viruses, bacteria, plants, and humans. Then they narrowed the focus to bacteriophages, viruses that infect only specific bacteria and are harmless to people.

How they picked the winners

The researchers didn't get it right on the first try, or even close. Out of 302 AI-generated genome designs that got synthesized and tested in the lab, only 16 actually worked well enough to kill E. coli.

Samuel King, a PhD student on Hie's team, described the moment they confirmed success. Bacteria-covered petri dishes started showing clear spots, meaning the new viruses were eating through the bacterial colonies. "We were starting to see these clear spots and it was just extremely exciting," King told the BBC. When the wider team saw the results, Hie says "the room spontaneously burst into applause."

According to Ars Technica, the viruses aren't some alien creation pulled purely from AI imagination. Every one of them is closely related to an existing natural virus. But they carry distinct features that would be difficult, maybe near-impossible, for evolution to produce on its own in the wild.

Why this matters for medicine

Antibiotic-resistant bacteria are a growing problem in hospitals, and phage therapy, using viruses to kill bacteria that antibiotics can't touch anymore, has been a niche but promising field for years. Hie argues the technology could "massively improve human health" by accelerating the design of new treatments that would take nature decades to stumble into.

This builds on earlier work: AI was already used to help design new antibiotic candidates targeting gonorrhea and MRSA, according to the BBC's own reporting from August 2025. Designing a functional virus genome from scratch is a significant jump up in complexity from designing a single protein or antibiotic compound.

The part that should make people nervous

The same AI approach that designed a harmless bacteria-killing virus could, in theory, be pointed at designing something that infects people.

The Stanford team says they built in a safeguard. Ars Technica reports the researchers deliberately excluded any viral sequences from complex cells, meaning anything that infects humans, animals, or plants, when they trained Evo1 and Evo2. The models were only fed data on viruses that go after bacteria and archaea.

That's a real precaution, not just a talking point. But it also means the safeguard is a design choice, not a hard technical wall. Nothing described in either source prevents a different team, with different intentions, from training a similar large genome model on human-infecting virus sequences instead. Dr. Thomas Inglesby and Dr. Moritz Hanke wrote a commentary accompanying the Stanford paper's publication in the journal Science, raising exactly this point: AI-designed viruses raise "urgent" safety and security concerns, not hypothetical ones down the road.

Biosecurity researchers have worried for years about "dual-use" research, science that's genuinely useful for medicine but could be repurposed to cause harm. This is that concern showing up in concrete, published, peer-reviewed form, not a theoretical worst-case scenario dreamed up by pundits.

At the same time, nothing in the BBC or Ars Technica reporting suggests anyone has actually built, or is close to building, an AI model capable of designing a virus that infects vertebrates. The researchers themselves say that's a future risk to prepare for, not a present capability. Ars Technica notes plainly: "This isn't science fiction" in the sense that these are real, working viruses, but it's also not evidence that a human-targeting version is around the corner.

What happens next

No government agency has announced new regulations on large genome models as of this writing. No legislation has been introduced specifically addressing AI-designed pathogens. The research was published openly in Science, meaning the methods are now public, which cuts both ways: it accelerates legitimate medical research and it means the underlying technique isn't secret.

The open question is whether biosecurity oversight, currently built around physical lab access and known pathogen lists, can adapt fast enough to a world where the dangerous part isn't a vial in a freezer but a training dataset and a few lines of code.