AI designs viable viral genomes, raising biotech promise and biosecurity fears

Scientists have used artificial intelligence to design viral genomes that do not exist in nature, and 16 of the resulting designs proved viable when they were synthesized and tested. The work, from researchers at the Arc Institute and Stanford University including Samuel King and Brian Hie, pushes AI from designing single genes to building whole genomes, a step that could reshape medicine and raise fresh biosecurity alarms.

The system was trained on more than 2 million bacteriophage genomes, then asked to generate candidate viral recipes. It produced 302 designs, and 16 of them worked. Several of the AI-designed phages lysed bacteria faster than the natural virus used as a template, a result that makes the achievement more than an abstract coding exercise. These were bacteriophages, viruses that infect bacteria rather than humans, so the testing stayed in bacterial systems rather than in people.

Arc Institute framed the project as a leap from individual genes to complete genomes, using genome foundation models from the Evo series. The preprint, titled Generative design of novel bacteriophages with genome language models, underscores how quickly the field is moving from protein sequences and gene fragments to larger biological architectures. That matters for public health because bacteriophage design could support antibacterial phage treatments, help scientists study how viruses evolve, and improve genome-scale research that feeds into vaccine and gene-therapy development.

The same capability also sharpens old worries about dual use. Once an AI system can generate functional viral material, the barrier to designing harmful biology can fall if the tools, data and synthesis pipeline are misused. That is why the policy debate has already widened beyond scientific novelty. RAND published a brief on October 1, 2025, asking when people should worry about AI being used to design a pathogen, and Nature ran a related warning that the world’s first AI-designed viruses marked a step toward AI-generated life. Testbiotech said on September 25, 2025, that AI-designed viruses had been created for the first time and pressed for caution.

What makes this moment newly urgent is not just that AI can now write plausible biological instructions, but that some of those instructions have already cleared the most basic reality check: they worked. Regulators, universities and biotech companies now face a harder task, deciding how to govern publication, access and oversight for a technology that could speed lifesaving science while also lowering the barrier to dangerous biological design.