AI designed viruses outpace nature — and US biosafety rules don’t cover them

In a lab at Stanford University, researchers watched clear spots spread across a petri dish full of bacteria — the unmistakable sign that something they had built entirely inside a computer had just come to life. The organisms wiping out that bacteria weren’t pulled from nature. They were AI designed viruses, generated by a computational model called Evo and then chemically built from scratch, and the results are now raising questions that go far beyond one Stanford basement lab.

Key takeaways

• Evo, an AI model developed by Stanford University and the Arc Institute, proposed 700,000 possible viral genomes, of which 285 were chemically synthesized and 16 produced viruses able to replicate and kill bacteria.
• Evo was trained first on roughly nine trillion nucleotides from millions of animals, plants, microbes and viruses, then specialized on the 11 genes of the bacteriophage Phi X-174 and about 15,000 related genomes.
• The AI-generated viruses were as robust as natural ones, and some replicated faster than the natural phage Phi X-174 itself.
• Current US National Institutes of Health biosafety rules ban experiments that make pathogens more dangerous, but they don’t clearly cover purely computational virus design unless it involves an “entity of concern.”
• Experts, including Moritz Hanke of the Johns Hopkins Center for Health Security, warn of a growing gap between what AI can now do and the rules meant to police it.

AI Model Evo Designs Novel Viral Genomes for Bacteria

Evo works like a genetic version of a large language model, predicting sequences of DNA the way ChatGPT predicts the next word in a sentence. Instead of language, though, it learned the patterns written into the genetic code of life itself. The team behind it, based at Stanford and the Arc Institute, set out to see whether that kind of pattern recognition could go a step further than designing individual proteins or antibiotics — and actually produce a working virus.

How Evo was trained

The model’s education happened in two stages. Evo’s initial training drew upon approximately nine trillion nucleotides sourced from millions of organisms spanning the animal, plant, microbial and viral kingdoms, absorbing patterns that run across the entire tree of life. Only after that broad foundation did a second, specialized round of training follow, this time focused narrowly on the 11 genes of the bacteriophage Phi X-174 and about 15,000 of its closest relatives. For Samuel King, a doctoral student at Stanford and co-author of the research, layering the training this way felt like the obvious move. “It just felt like the obvious next step,” he said.

From 700,000 candidates to 16 working viruses

Once trained, Evo proposed 700,000 possible viral genomes. The researchers narrowed that enormous pool down to the most promising candidates and had 285 sequences chemically synthesized as DNA, which were then inserted into bacteria to see what would happen. Sixteen of those sequences produced viruses capable of replicating — and, crucially, capable of killing the target bacteria, E. coli, in the process. King and his colleagues reportedly realized the phage were working in the early hours of the morning, watching clear spots open up on the petri dishes. “We were starting to see these clear spots and it was just extremely exciting,” King said. When the wider team saw the results, according to Stanford assistant professor Brian Hie, “the room spontaneously burst into applause.”

How AI-Designed Viruses Actually Performed in the Lab

The 16 viruses that emerged from the experiment weren’t fragile lab curiosities. They matched the performance of naturally occurring viruses, and in some cases outdid them — a result that surprised even researchers not involved in the work.

Robustness, speed and scientific milestones

The designed viruses proved as robust as natural ones, and some replicated even faster than Phi X-174 itself. “They’re not just sickly versions of stuff that already exists,” said Oliver Crook, a protein chemist at the University of Oxford who wasn’t part of the study. Patrick Cai, a synthetic biologist at the University of Manchester, called the work an “important milestone,” adding that “the significance extends far beyond phages — it suggests that genome language models are beginning to learn the design principles encoded by evolution, opening the door to AI-assisted genome writing.” Marc Güell, from the synthetic biology lab at Pompeu Fabra University, described it as a “very significant turning point” because, for the first time, “we are beginning to design biology on a computer.”

Still, Crook was careful to temper expectations. The AI didn’t invent anything fundamentally new — the resulting viruses are very similar to natural species and rely on the same underlying biology rather than some novel mechanism. Whether Evo could pull off the same trick with other virus families remains an open question. If the approach does generalize, Crook said, it could yield genuinely useful tools for medicine and biotechnology: “A lot of our science rests on viruses as technology.” For context, the phage genome the team worked with runs about 5,400 base pairs, compared with roughly 500,000 base pairs in the smallest living cell and around three billion in the human genome — a gap Hie says makes designing a simple living organism “probably a lot of work, but not impossible,” and something the team is “definitely interested in working towards.”

Biosafety Gaps Around Computational Virus Design

The breakthrough lands at a moment when US biosafety policy hasn’t caught up with what AI models can now generate on a screen. That mismatch is at the center of the debate around virus genome synthesis guided by artificial intelligence.

What US biosafety rules don’t cover

The In late July, the National Institutes of Health unveiled a policy regarding high-risk life sciences research that prohibits the creation of more hazardous pathogenic variants. However, work conducted entirely through computational methods — designing viral DNA on a computer without physically building it — isn’t covered “unless it involves an entity of concern,” according to the agency. That distinction is easy to apply to something like smallpox, which is clearly classified as dangerous. It’s far murkier when the virus in question came out of an AI model rather than a known pathogen list, which is exactly the kind of ambiguity now shaping debates over biosafety regulation AI policy struggles to keep pace with.

Precautions and expert warnings

The Stanford and Arc Institute team didn’t wait for regulators to catch up. Throughout its training phase, Evo was not exposed to any information concerning human-infecting viruses or their associated pathogenic organisms affecting animals, plants or fungi — meaning the model simply cannot produce those genetic sequences to begin with. “Our intention was simply to exercise maximum caution,” said Brian Hie. Moritz Hanke of the Johns Hopkins Center for Health Security called that decision “quite commendable,” especially since no official rules required it: “Because they don’t get any guidance from anywhere on what they should be doing.”

Hanke and Thomas Inglesby, also of Johns Hopkins, went further in a commentary published alongside the research in the journal Science, warning the findings raise “urgent biosafety and biosecurity questions.” The real issue, they argued, isn’t whether generative viral genome design will exist going forward — it clearly already does — but whether it can be used without “enabling serious harm.” Hanke laid out the concern bluntly: “What is the risk of what I’ve never seen before?” he asked, describing “a huge disconnect” between how fast the science is moving and how slowly the guardrails are following. His worst-case scenario is straightforward to imagine: “You could say, ‘Hey, genomic language model, make me an influenza genome that is modified to be more transmissible or to be more lethal.'”

Why this matters: the same computational shortcut that let Evo design 16 working bacteriophages in a matter of months, using nine trillion nucleotides of training data and a targeted follow-up round on Phi X-174, is a general-purpose method. Nothing about the underlying technique restricts it to harmless bacteria-killers. That’s precisely the gap Hanke and Inglesby are pointing to: a research method that’s advancing faster than the policy meant to contain it, in a field where synthetic biology research increasingly blurs the line between designing new medicines and designing new risks.

For now, the researchers say the safeguards they applied — excluding human-infecting virus data, working only on bacteriophages, and running everything inside a secure lab — go a long way toward keeping the technology pointed at benefit rather than harm. Hie argues the approach has the potential to “massively improve human health” through new drugs and therapies, including phage-based treatments for antibiotic-resistant infections. Whether that promise holds up outside Phi X-174, and whether regulators move before the next version of this experiment targets a less contained target, are the two questions likely to define what comes next for AI-designed viruses.

FAQ

How effective were the AI-designed viruses in laboratory tests?

Out of 285 synthesized viral genome sequences, 16 produced viruses capable of replicating and killing bacteria in the lab.

What data was used to train the AI model Evo?

Evo was first trained on roughly nine trillion nucleotides from millions of animals, plants, microbes, and viruses, then specifically on the 11 genes of phage Phi X-174 and about 15,000 related genomes.

Why do current US biosafety policies not regulate AI-designed viruses?

Because purely computational viral DNA design is not covered under current US NIH policies unless it involves an entity of concern, creating a regulatory gap for AI-designed viruses.

What precautions did the scientists take during Evo’s training?

They excluded data on viruses that infect humans and related pathogens, to prevent the AI from generating potentially harmful human-infecting viral genomes.

Article produced with the assistance of artificial intelligence and reviewed by the editorial team.