AI designs living viruses never seen in nature

Language models aren't just writing Python anymore — they're writing functioning genomes from scratch.

Researchers from Stanford and the Arc Institute just used AI to design 16 viruses that do not exist in nature. Published in the journal Science, this marks the first time a language model has generated complete, fully operational genomes that came to life in a petri dish.

From code prompts to living phages

The team worked with Evo 1 and Evo 2, models trained on millions of evolutionary genomes. They tasked the AI with rewriting Phi X174, a textbook virus that specifically targets and infects E. coli bacteria.

Out of 285 AI-designed phages synthesized and tested in the lab, 16 were viable. Some of these synthetic viruses actually replicated faster than the wild-type original, and a few differed enough from known genomes to qualify as entirely new species.

Crucially, when researchers faced E. coli strains that had grown resistant to the natural virus, a cocktail of these AI-engineered phages wiped them out completely. It serves as a striking proof of concept for personalized therapies targeting antibiotic-resistant bacterial infections.

On the biosafety front, the researchers noted they strictly avoided training the models on viruses that infect humans, animals, or plants, ensuring the system lacked the biological blueprints to threaten people.

Why it matters

We think this is one of the clearest demonstrations yet of generative AI solving a physical, high-stakes medical bottleneck. Superbugs and antibiotic resistance are slowly becoming one of modern healthcare's biggest nightmares. Being able to prompt an AI for custom phages that outmaneuver bacterial mutations could compress decades of drug discovery into weeks.

However, the biosecurity implications are impossible to ignore. The core mechanism used to neutralize bacteria is functionally identical to what someone would use to engineer a human pathogen. And because Evo 2 is open source, the underlying weights and architectures are out in the wild. Stanford and Arc responsibly sanitized their training data, but nothing stops bad actors from fine-tuning similar models on toxic viral agents.

In our view, biology is having its dual-use software moment. Just as vulnerability research tools can patch zero-days or fuel ransomware, AI genome generators can save lives or unleash hazards. The real battle won't just be inside the weights, but at the DNA synthesis screening facilities responsible for printing these digital files into physical reality.