AI Designed 16 New Viruses That Actually Work—and That's Both Good and Terrifying

For the first time, scientists have used generative AI to design whole viral genomes from scratch—and then watched those viruses come to life in a petri dish. Researchers at Stanford University and the Arc Institute trained an AI model called Evo on millions of genetic sequences from across the tree of life, then asked it to propose completely new genomes for bacteriophages—viruses that infect bacteria, not humans. Of the roughly 300 AI-generated designs they synthesized and tested, 16 turned out to be fully functional, capable of infecting and killing E. coli bacteria. The results, published in Science on Thursday, mark what multiple reports describe as the first demonstration of generative AI producing entire viral genomes that can replicate in the lab.

The Evo model, trained on a massive corpus of genetic data, churned out thousands of candidate phage genomes. The team selected a subset of about 300 for physical synthesis and laboratory testing. Sixteen of those—roughly 5 percent—proved viable, meaning the viruses could successfully infect E. coli and complete their life cycle. While a 5 percent success rate may sound low, the fact that any of these wholly AI-proposed sequences worked at all is a significant scientific milestone. Previous efforts have focused on editing existing viruses or using AI to predict protein structures; this is the first time a generative model has proposed an entire viral genome that then functions in the real world.

Importantly, the viruses created are bacteriophages, which pose no direct threat to humans. Several reports explicitly state these phages cannot infect human cells. The team also found that a mixture of the AI-designed phages could overcome antibacterial resistance in some E. coli strains more effectively than a comparable cocktail of naturally occurring phages. This hints at a practical application: using AI to engineer phage therapies that are optimized to defeat drug-resistant bacteria.

The study lands at a moment when antibiotic resistance is accelerating worldwide, and the pipeline for new antibiotics has largely dried up. Phage therapy—using viruses to kill specific bacteria—has long been seen as a promising alternative, but traditional methods of finding and characterizing phages are slow and labor-intensive. AI could dramatically speed up the process, generating thousands of candidate phages tailored to target specific bacterial strains.

Beyond medicine, the work demonstrates that generative AI can navigate the immense complexity of biological sequence space. The Evo model was trained on data from millions of genomes across all domains of life, learning the statistical patterns that make a sequence biologically plausible. The fact that it can now propose functional viral genomes suggests the same approach could be applied to other biological systems—designing new enzymes, metabolic pathways, or even minimal cells.

Yet the same capability that makes this exciting for drug discovery also raises alarms. Experts quoted in the coverage warn that if similar AI tools were redirected toward dangerous human pathogens, they could lower the barrier to creating novel bioweapons. The study's authors and commentators are careful to note that the current work involves only bacteriophages and that no human-infecting viruses were created. But the underlying technique—generative AI for whole-genome design—is pathogen-agnostic.

For the biotech industry, this development signals a shift in what's possible with AI-driven drug discovery. Startups focused on synthetic biology and phage therapy—such as Locus Biosciences, Eligo Bioscience, and others—may find themselves with a new competitive tool. Large pharma companies investing in antimicrobial resistance programs could also take note: AI-designed phages might offer a faster, cheaper path to clinical candidates.

However, the regulatory landscape is uncertain. If AI can generate novel viral genomes on demand, regulators will need to decide how to evaluate the safety and security of such products. Companies working in this space should expect increased scrutiny and may need to invest in biosecurity protocols from the outset. The research also underscores the importance of open science: the Evo model and its training data are likely to be shared, but the question of who gets to use such models—and for what purposes—will become a pressing business and policy issue.

The immediate next step will be scaling up: can the success rate of viable designs be improved from 5 percent to something closer to 50 percent? More importantly, can the approach be extended to design phages that target the most dangerous antibiotic-resistant pathogens, such as MRSA or C. difficile? Watch for follow-up studies that test AI-designed phages in animal models, and for regulatory moves from bodies like the NIH or the WHO on guidelines for AI-driven pathogen design. The line between breakthrough and breakthrough risk has never been thinner.