What happened.

A team from Stanford and Arc Institute used an AI model called Evo to design complete viral genomes from scratch. Out of 285 sequences synthesized in the lab as DNA and inserted into E. coli bacteria, 16 produced functional viruses capable of replicating and destroying their targets. The work is published in Science (August 7, 2026). Evo had proposed 700,000 possible genomes; the team synthesized the most promising candidates. Some of the generated viruses replicate faster than the natural phage Phi X-174 used as a reference.

Why it matters to you.

AI agents so far worked in text and code. Here a model generated something that leaves the computer and does a measurable thing in the physical world. As we reported in July on scientific computing, agentic AI is entering laboratories. The leap is of a different order: the model directly writes the biological sequence, and the lab verifies it.

In detail

Evo works like a language model, but trained on DNA instead of text. The first training phase consumed approximately nine trillion nucleotides from millions of animals, plants, microbes and viruses, learning patterns that span the entire tree of life. The second phase specialized in the phage Phi X-174, a simple virus with 11 genes and approximately 5,000 letters of DNA, and 15,000 of its closest relatives.

The experiment numbers. Evo proposed 700,000 genomes. The team had 285 synthesized as DNA and inserted them into E. coli bacteria. Sixteen produced viruses capable of replicating. These are small numbers in absolute terms, but sufficient to demonstrate one thing: the model produces genomes that function in a living cell. On paper they seemed plausible; in the bacterium they replicated.

What external reviewers say. Oliver Crook, protein chemist at Oxford University, described the viruses as robust as natural ones: “They are not sickly versions of things that already exist.” Patrick Cai, synthetic biologist at Manchester, speaks of “an important milestone.” Both temper the enthusiasm. The generated viruses are very similar to natural species and based on the same biology. Evo didn’t invent anything fundamentally new. It remains open whether the model would be equally effective with other viral groups.

The regulatory gap. The National Institutes of Health published a policy on high-risk research in late July that prohibits experiments that make pathogens more dangerous. Purely computational work, that is designing viral DNA on a computer, does not fall under the policy “unless it involves an entity of concern.” With smallpox the classification is clear. With a virus generated by an AI model it is not. Moritz Hanke of Johns Hopkins Center for Health Security puts it plainly: there is an “enormous gap” between the pace of research and the guardrails surrounding it.

The precaution taken voluntarily. During training, Evo did not receive data on viruses that infect humans, nor on related pathogens of animals, plants or fungi. The model cannot generate those genomes in the first place. Brian Hie, computational biologist at Stanford and co-author, calls it a choice of prudence: “We wanted to be extra-cautious.” No official rule required it.

What changes and what doesn’t. The work opens a path toward tools for medicine and biotechnology: viruses that kill bacteria have real applications, from alternatives to antibiotics to basic research. The leap from bacteriophages, which infect bacteria, to viruses that infect humans is enormous and deliberately excluded from the model’s design. The most concrete result today is methodological: a generative model can produce entire verifiable genomes, and the physical laboratory remains the necessary validation phase.