AI designs first working viruses – 16 bacteriophages built by computer

For the first time, artificial intelligence has designed complete, functioning blueprints for a life form: a team from Stanford and the Arc Institute had AI models write full genomes of bacteriophages – viruses that infect bacteria. The results appeared on 6 August 2026 in the journal Science.

The template was ΦX174, a harmless, well-studied bacteriophage with about 5,386 nucleotides across eleven genes. The genome language models Evo 1 and Evo 2 generated roughly 700,000 candidate genomes. Researchers synthesized 285 designs – and 16 of them actually worked: they infected and killed E. coli bacteria. Three even outperformed natural strains. A success rate of about 5.6 percent.

The scientific appeal is enormous: tailor-made phages could one day be deployed against antibiotic-resistant bacteria – a multi-billion-dollar problem in medicine. That an AI can design working genomes from scratch was science fiction until recently.

But the breakthrough raises hard questions. Researchers at Johns Hopkins University warned bluntly: 'The ability to compose viral genomes using generative AI now exists; the governance to safely steer it does not.' It is also notable that the recent US gain-of-function policy explicitly exempts computational design work.

Other voices push back. Tom Ellis of Imperial College London considers the danger overstated: modifying existing pathogens remains much easier than designing viruses entirely from scratch. The researchers also stress that they worked only with bacteria-infecting phages – not with pathogens dangerous to humans.

For the AI world, August 2026 marks a turning point: generative models are leaving the screen and reaching into biology. The technology is here. The rules, as the journal team itself writes, urgently need to catch up.