AI Just Designed Living Viruses That Never Existed in Nature
Stanford and Arc Institute's Evo genome models designed novel bacteriophages that replicate in E. coli. Science paper: 700k designs, 285 built, 16 worked - some beating ΦX174. Biosecurity alarms and phage therapy hopes collide.
Today’s Science paper is the kind of headline that makes people check the date.
Researchers at Stanford and the Arc Institute used genome-scale language models - Evo 1 and Evo 2 - to design novel bacteriophages: viruses that infect bacteria, not humans, with DNA sequences that did not exist in nature until a model proposed them and a lab built them.
They generated on the order of 700,000 designs. They synthesized 285. Sixteen produced working phages that could replicate in E. coli. Some outperformed the classic workhorse phage ΦX174 on key lab measures. That is not a demo video with clever editing. That is wet-lab reality: AI proposed genomes, humans ordered DNA, viruses lived.
What Evo actually is
Think of Evo less like ChatGPT for essays and more like a model that learned the grammar of genomes - patterns across DNA that correlate with functional biology. Evo 2 is open, which thrills researchers and unsettles biosecurity people in the same breath. The team says training voluntarily excluded human and animal virus genomes - a deliberate fence, not a law of physics. Fences can be rebuilt. They can also be ignored by someone else with compute and less restraint.
The scientific upside is obvious if you work in microbiology. Designing phages by hand is slow. Antibiotic resistance is not. If models can propose phages that kill specific bacteria more effectively, phage therapy - using viruses as precision antibiotics - gets a jet engine. Hospitals fighting drug-resistant infections have been waiting for that sentence for years.
The part that keeps people up
Tom Inglesby and Kevin Hanke, among other biosecurity voices, are already in the worry column. The concern is not that these sixteen phages will hunt humans. They are bacteriophages. The concern is the capability: a pipeline that can invent functional viral genomes at scale, synthesize a subset, and find winners. Today the target is E. coli. Tomorrow the same methods, with different data and fewer scruples, could aim elsewhere.
Open weights for Evo 2 make the debate sharper. Open science accelerates medicine and also lowers the barrier for misuse. Closed models concentrate power in labs and companies that may or may not be careful. There is no free lunch - only tradeoffs with body counts attached if we get them wrong.
How big is “16 worked”?
In biotech, a low hit rate can still be a breakthrough if the hits are real. 16 / 285 is roughly 5.6%. For novel genomes dreamed up by a model, that is extraordinary. Nature spent billions of years on trial and error. A cluster of GPUs and a synthesis budget compressed a slice of that search into a project timeline.
Some of the winners beat ΦX174 - a phage so well studied it is practically a lab mascot. Beating the mascot is how you know you are not just rediscovering textbook sequences with fancy branding.
What responsible looks like (and what it does not)
Voluntary exclusion of vertebrate virus genomes is good practice. It is not a global safeguard. Synthesis screening, lab access controls, publication norms, and export rules for nucleic acids matter more the day after a Science paper than the day before - because suddenly everyone knows the trick works.
Phage therapy advocates will celebrate. Dual-use analysts will draft memos. Policymakers will ask whether “AI designed a virus” needs a new law or just better enforcement of old ones. The honest answer is probably both, plus funding for the boring infrastructure: screening orders, red-team evaluations, and international norms that do not assume every actor shares Stanford’s ethics memo.
The magazine version of the stakes
We used to argue about AI writing essays and faking voices. That argument feels quaint next to AI proposing genomes that assemble into replicating particles. The particles in this paper eat bacteria. That is genuinely useful. It is also a proof of concept that biology’s code is becoming a generative medium.
If you want the optimistic frame: we might get better tools against superbugs. If you want the pessimistic frame: we just taught software to brainstorm life that can copy itself, and we published the syllabus.
Both frames are true on August 7, 2026. The Science paper does not choose for you. Policy will have to.