Sixteen new viruses now exist that have never existed before, and an AI model made them. That’s not a hypothetical. That’s what happened.
As reported by Engadget, a study published in the journal Science details how researchers at the Arc Institute in Palo Alto and Stanford University used genome language models called Evo 1 and Evo 2 to design entirely new viruses capable of infecting and reproducing inside bacteria. The models work on roughly the same principle as large language models like ChatGPT, except instead of training on text, they trained on genetic sequences. Trillions of nucleotides. The building blocks of DNA. The team taught these systems the grammar of genetic code, and then asked them to write something new.
For the experiment, the researchers focused on about 15,000 viruses from the same family as Phi X-174, a small virus that only infects E. coli bacteria. They explicitly limited the scope to exclude anything capable of infecting humans, animals, plants, or fungi. Evo generated roughly 700,000 possible new viruses. The team narrowed that down to 285 candidates, manufactured their DNA, and inserted it into bacteria. Sixteen of those candidates produced working, viable viruses. Some reproduced faster than the original Phi X-174.
The researchers were careful. They set limits. They published their findings. And the potential upside is real. AI-designed viruses could improve gene therapies, give scientists better tools for targeting harmful bacteria, and help researchers understand how genomes actually work. In a tightly controlled setting, with real ethical oversight, this is the kind of science that eventually helps people.
But here’s where the privacy and security angle becomes impossible to ignore. We already know that standard chatbots, the kind anyone can access right now, have provided users with concerning assistance in thinking through biological threats. A model that doesn’t just describe biology but actively designs functional genetic sequences is a different category of risk entirely. The step from “explaining how viruses work” to “generating a working virus” is not a small one.
The regulatory picture around AI biosecurity is not keeping pace with the science. It isn’t close. And while these particular researchers acted responsibly, the models and the methods they used are not locked away. The knowledge that this approach works is now public.
So what does this mean practically? A few things worth watching:
- Genome AI models could become tools for designing targeted bacteriophages to fight antibiotic-resistant infections
- The same models could, in less careful hands, be used to explore dangerous pathogens
- Existing biosecurity frameworks were not built with AI-assisted virus design in mind
- No clear international consensus exists yet on how to regulate this kind of research
The science itself is not the problem. But when a system can move from pattern recognition to functional biological creation, hoping that only responsible actors use it is not a strategy. It’s a wish. And wishes don’t scale.