AI writes living viruses: first genome designs work, and now a patient trial

What this is about

Researchers at Stanford University and the Arc Institute in California have used artificial intelligence to design complete genomes of working viruses for the first time. The study appeared in the journal Science in August. The viruses are bacteriophages (phages for short), which infect only bacteria. PhD student Samuel King told the BBC that he saw clear spots in a lawn of bacteria on Petri dishes at night: a virus designed on a computer had killed the bacteria there. According to lab head Brian Hie, the room spontaneously broke into applause.

The story has moved on since then. According to the New York Times (cited by ProPakistani), AI-designed viruses have now been used on a human patient for the first time. At the same time, experts are debating whether biosecurity is keeping pace with the technology.

How the AI designed the viruses

The template was ΦX174, a harmless phage that has been studied for decades and attacks only one strain of gut bacteria. Its genome is only about 5,400 DNA letters long and holds the instructions for just 11 proteins. That makes it a simple test system with a clear-cut outcome: the virus either works or it doesn't.

Evo 1 and Evo 2 are so-called genome language models. They learned the "language of DNA" much as ChatGPT learned the language of words: by reading huge numbers of real examples. The team trained them on more than two million phage genomes and fine-tuned them on about 15,000 viruses closely related to ΦX174. According to the sources, Evo 2 had seen genomes from across the living world, but viruses that infect complex organisms were deliberately left out. According to The Conversation, the model accordingly performed poorly on proteins from human viruses.

According to ProPakistani, the model generated around 700,000 genomes, from which 285 were selected for lab testing. 16 produced viable viruses.

Results and expert criticism

In a six-stage pipeline, most candidates were filtered out based on sequence quality, matching bacterial host and distance from natural sequences. The share of working designs is small: the author of the War on the Rocks piece puts it at around 5 percent. Systems biologist Jordi García Ojalvo said so much was wrong that the designs could be seen as "hallucinations" of the model. Simon Jackson (University of Waikato) pointed out that half of the 16 working viruses picked up mutations along the way, a sign of natural evolution at work and not just of perfect AI designs.

Still, the results are real: some viruses killed bacteria faster than the natural ΦX174. Simon Clarke (University of Reading) sees significance here for phage therapy against antibiotic-resistant germs. Patrick Cai (Manchester) says the models appear to be learning evolution's design rules, which opens the door to AI writing other genomes. According to War on the Rocks, only designs with at least 92 percent identity to an existing virus survived. In other words, the AI remixed what was already known instead of inventing something entirely new.

First use in a human

Three people from the project, including Eric Nguyen, founded the company Radical Numerics. In June 2026 it unveiled the model Omnii, which processes DNA sequences, protein structures and other biological data. Omnii was used to design variants of the phage T7, with altered "leg structures" so that they dock onto E. coli strains that earlier phages could not attack. The DNA was synthesized, introduced into bacteria and tested. After safety checks, the viruses went to the treating doctors. The US FDA approved the experimental use. On October 20, the company gave no details about the patient or doctors and did not say whether the infection was beaten. The results have not been peer-reviewed. Nguyen hopes to create personalized treatments within days in the future, instead of searching for matching phages for months.

Biosecurity: how big is the risk?

Several experts say the leap from a bacterial virus to a threat to humans is enormous. For comparison: the phage genome has about 5,400 letters, the smallest genome of a living cell about 500,000, and the human genome three billion. Martin Täuber (Zurich) compared it to nuclear bomb blueprints: the plan isn't the problem, the uranium is. The real bottleneck is materials, equipment and lab work. Jason Kelly (Ginkgo Bioworks) said that people in the lab can refuse requests from an AI. A study mentioned by War on the Rocks had more than a hundred students and others attempt four tasks involving reviving a virus. Language models such as ChatGPT-5 did not measurably improve overall success.

Experts still urge caution. Scientists at the Johns Hopkins Center for Health Security (including Tom Inglesby) wrote in a Science commentary that the open question is no longer whether AI can design viral genomes, but whether this can be done without serious harm. New models should be independently evaluated before release. Some synthetic DNA providers screen orders with software, but no law requires it. According to War on the Rocks, an open letter calls for mandatory screening and documentation of DNA orders. Author Jason Kaelber warns that AI can already alter properties of existing viruses and that the teams needed are getting smaller. The Conversation stresses that the study does not show AI can design a pandemic virus, but that complete, working viral genomes are feasible. Evo 2 has also been released openly, which makes control harder.

Our take

The work is a genuine milestone, but not a doomsday scenario: the target was harmless, the hit rate was low, and lab work remains indispensable. Medically, the path to tailored phages against resistant germs looks promising, but the clinical benefit is unproven. The first patient case comes with no outcome and no peer review. It remains open whether safeguards such as training-data filters, DNA screening and monitoring will become binding in time.