Generative AI has crossed a new threshold in biology: designing a complete, functional viral genome from scratch, rather than simply analyzing or tweaking one that already exists. Researchers at Stanford University and the Arc Institute, publishing in the journal Science on August 6, 2026, used AI models called Evo 1 and Evo 2 to generate entire bacteriophage genomes — viruses that infect bacteria, not humans — capable of assembling into working virus particles and replicating inside living cells. Led by Stanford's Brian Hie, the team describes it as a genuinely new level of complexity for generative biology, well beyond earlier AI work that focused on individual proteins or candidate antibiotics.

The Evo models work by learning the underlying patterns of DNA the way a language model learns grammar, rather than being told explicit rules. Evo 2, the larger of the two, was trained on roughly 9.3 trillion base pairs drawn from around 128,000 genomes spanning bacteria, viruses, plants, and animals. Using a well-studied E. coli-infecting virus as a structural template, the team synthesized 302 AI-designed genome candidates in the lab; 16 of them proved fully functional, successfully infecting and killing E. coli, with some reportedly performing on par with or better than their natural counterparts. To limit risk, the researchers deliberately excluded any virus capable of infecting humans or other complex organisms from the training data, and the work was carried out entirely under secure lab conditions targeting only bacteria.

Even so, outside experts are quick to note how far this still is from designing life at large. A bacteriophage genome runs to about 5,400 DNA base pairs; the simplest living cell needs roughly 500,000, and the human genome spans billions. Researchers not involved in the study, including a synthetic genome specialist at Imperial College London, have pointed out that phage genomes are unusually small and tolerant of mutation compared to more complex organisms, meaning the difficulty of this kind of design likely scales exponentially rather than linearly as genome size grows. The immediate promise is narrower and more practical: engineered phages are already being explored as a therapy against antibiotic-resistant bacterial infections, a problem conventional drug discovery has struggled to keep pace with.

The harder conversation is playing out alongside the science itself. A companion editorial in the same issue of Science, from Johns Hopkins biosecurity researchers, credits the Stanford team for building in real safeguards but argues the governance needed to safely steer this capability more broadly doesn't yet exist — and frames the open question as no longer whether AI-driven genome design will happen, but how oversight catches up to it. Biosecurity researchers elsewhere have echoed the concern: the same generative approach that designs a helpful, narrowly-targeted phage could, in principle, be redirected toward more dangerous ends if applied outside a controlled research setting. Current biotechnology regulation is described by multiple outlets as a patchwork that hasn't yet caught up to AI-generated genomes specifically, even as recent policy moves have tightened rules around high-risk "gain of function" research more broadly. For now, the achievement stands as a genuine first — and a preview of a debate that's only getting started.