Scientists have used an artificial intelligence program called Evo to design new viral genomes that are different from any known natural viruses — and 16 of them turned out to be viable, living viruses when built and tested in the lab.
The study, published Thursday in the journal Science and detailed in CNN's reporting, is being hailed as a milestone for generative biology. But it is also drawing urgent warnings about biosecurity, because it demonstrates that the same techniques powering chatbots can now compose functional, novel pathogens. For more context on this story, see our ongoing AI trends.
How Evo Was Trained
According to CNN, genetic sequences drawn from millions of sources — "from all domains of life" — served as the language library for Evo, much like the way ChatGPT and other large language models are trained on huge collections of text. By studying that library, the model was able to "learn the evolutionary constraints" that govern natural genomes — the unwritten rules about which genetic arrangements actually work in living things.
Scientists from Stanford University and the Arc Institute then directed Evo to generate thousands of genome combinations within a specific framework — in this case, genomes that could be hosted by an E. coli bacterial cell. Of the thousands of designs the model produced, the team built and tested roughly 300 in the laboratory. Sixteen of them were viable viruses.
Bacteriophages, Not Human Pathogens
An important caveat tempers the alarm: the new viruses are bacteriophages, which infect bacteria but cannot infect humans. The work therefore falls on the hopeful side of the technology's potential. The researchers reported that one of the new viruses carried an aspect that was "evolutionarily distant," suggesting Evo had prompted genetic changes that could have taken millions of years of natural evolution to emerge — compressing deep evolutionary time into a single generative step.
The most striking practical result came from mixing the AI-designed phages together. Tests showed the mixture could overcome antibacterial resistance in some E. coli strains — something a "comparable mixture of naturally sourced" phages could not do. That is no small claim. Antibiotic resistance is one of the gravest creeping threats in global health, and phage therapy — using viruses to kill harmful bacteria — has long been pursued as a possible answer to drugs that no longer work.
The authors wrote that the approach "lays out a path for generating adaptive and resilient phage therapies against rapidly evolving pathogens," framing Evo as a design engine for medicines that could keep pace with the bacteria they are meant to defeat.
"Urgent" Biosafety Questions
The promise comes paired with concern that the researchers themselves acknowledge. In a corresponding article also published in Science, doctors from the Johns Hopkins Center for Health Security struck a cautionary tone. "Although this is promising for life sciences applications, it also raises urgent biosafety and biosecurity questions," they wrote. "The ability to compose viral genomes using generative AI now exists; the governance to safely steer it does not."
They warned that such genomes "might encode new pathogens that can infect humans, animals, or plants in ways that cannot be contained by existing countermeasures" — a reminder that the same generative method that produced harmless bacteriophages aimed at E. coli could, in principle, be pointed at other targets.
Not every expert considers the immediate risk extreme. Jordi García Ojalvo, a professor of systems biology at the Pompeu Fabra University of Barcelona, told the Science Media Centre that "the breakthrough achieved is significant." CNN noted his assessment that the biosafety risk posed by this particular research is lower than for some other AI tools, largely because the genomes still have to be individually built and tested in a lab after Evo designs them — a physical bottleneck that does not exist for text or images.
Why It Matters Beyond Biology
The Evo result is the clearest demonstration yet that generative AI has crossed from language and images into the design of functional biological entities. It echoes the tension that now defines much of the AI field: the same capabilities that promise new medicines and resilient therapies also lower the effort required to design potentially harmful organisms.
The Johns Hopkins commentators argued that the research community needs to "engage with biosafety and biosecurity questions more deliberately than most developers of powerful biological AI models" have so far. That argument applies well beyond Evo — it is a warning about an entire emerging category of biological AI systems whose outputs can, unlike a paragraph or a picture, be physically instantiated in the world.
What Comes Next
The Stanford and Arc Institute team has framed the work as a foundation for phage-based therapeutics against drug-resistant bacteria. The harder, unresolved challenge is governance: how to release and oversee biological AI tools whose designs can be synthesized into living organisms. As the twin Science papers make clear, that is now a live question rather than a hypothetical one — and the gap between capability and oversight is exactly what the field's most cautious voices are now racing to close.
---
Stay Ahead of AIGet the latest AI news, analysis, and breakthroughs — all in one place.
Read more AI news →
