For a few scientists in a Stanford laboratory, the first sign that something remarkable had happened was not a headline or a press conference.
It was applause.
When the results from their experiment were presented to the lab, Stanford researcher Brian Hie recalled to NPR that “the lab spontaneously broke into applause,” an unusual reaction for a research meeting. The reason was equally unusual: artificial intelligence had helped produce viral genomes that worked in the real world.
The experiment, published in Science in August, involved bacteriophages, viruses that infect bacteria. Sixteen AI-designed phages proved viable in laboratory testing, marking a significant step in the use of generative AI for whole-genome design.
The achievement is exciting for scientists interested in new ways to fight difficult bacterial infections.
It is also the reason biosecurity experts are asking a harder question: if AI can help design a complete viral genome today, how should the world prepare for what these systems might be able to design tomorrow?
AI Is No Longer Just Reading Biology
The technology behind the experiment works in a way that will sound familiar to anyone who has used a large language model.
Instead of learning patterns in books and websites, genome language models learn patterns in DNA. The models used in the study, Evo 1 and Evo 2, were developed to work with genetic sequences, allowing them to generate biological code rather than ordinary text.
The difference is profound.
A chatbot can produce a sentence that makes little sense and the mistake is usually obvious. A genome sequence is different. Whether a generated sequence actually functions cannot be settled by simply reading it. It has to be tested experimentally.
That is what makes this study important.
The researchers were not merely asking an AI system to predict what a gene might do. They were testing whether it could help design an entire viral genome capable of functioning inside a biological system.
Sixteen Designs Worked. That Number Matters
The researchers generated a large number of candidate genomes and selected nearly 300 for laboratory testing. Only 16 proved viable.
That low success rate is an important part of the story because it shows how far the technology still has to go. AI did not simply produce hundreds of perfect viruses on demand.
Most of the designs did not work.
But the 16 that did were enough to demonstrate that the approach could cross the gap between a digital sequence and a functioning biological system. The peer-reviewed study describes the resulting phages as having diverse fitness profiles, while some experiments showed generated phages performing better than the natural template.
That is the threshold that has caught the attention of biosecurity researchers.
Then Comes the Biosecurity Problem
The same capability that could help scientists search for useful biological designs can raise questions about misuse.
Thomas Inglesby and Moritz Hanke of the Johns Hopkins Center for Health Security published a companion Perspective in Science alongside the study. They acknowledged the potential benefits while warning that the emergence of functional AI-designed viral genomes creates a governance challenge.
“The ability to compose viral genomes using generative AI now exists; the governance to safely steer it does not,” they wrote.
That warning needs context. The experts were not saying that the Stanford experiment had created a dangerous pathogen. Their concern is what happens if increasingly capable biological AI becomes easier to access and can eventually be applied to more complicated organisms.
Today, turning a digital design into physical biology still requires specialized expertise, laboratory infrastructure and access to DNA synthesis.
Those barriers matter. But technology has a history of making difficult tasks easier, cheaper and more accessible over time.
The Rules Are Moving, But They Are Not Finished
The United States already has a framework for nucleic acid synthesis screening, but the policy landscape is changing.
The U.S. Administration for Strategic Preparedness and Response says federal agencies are revising or replacing the 2024 framework following a 2025 executive order. The current government page says it will be updated when the new framework becomes available.
Congress has also considered legislation aimed at strengthening nucleic acid synthesis security. The Biosecurity Modernization and Innovation Act, for example, was introduced in the House in August 2026 and referred to the Energy and Commerce Committee.
The policy challenge is therefore not simply about regulating AI. It is about building several layers of protection around a technology that sits between computing, biotechnology and physical manufacturing.
This Is Bigger Than 16 Viruses
The significance of the study is easy to misunderstand if the number 16 becomes the entire story.
Sixteen functioning bacteriophages is not evidence that AI can now manufacture any virus scientists can imagine. The experiment was conducted on a small, well-understood biological system, and substantial technical barriers remain.
But the experiment does establish something that was previously much harder to demonstrate.
AI can help design a complete viral genome that works.
That changes the conversation from “Could this eventually happen?” to “How should we manage a capability that has now been demonstrated?”
And that is why the debate extends well beyond the Stanford laboratory.
The Next Question Is Not Whether AI Can Design Biology
It is how responsibly humans will use that ability.
The researchers have shown one possible future, where AI helps scientists explore biological designs that evolution may never have produced and potentially opens new approaches to difficult medical problems.
Biosecurity experts are looking at the other possibility, where increasingly capable systems reduce some of the expertise and time required to design biological systems that could be misused.
Both possibilities can be true at the same time.
That is what makes this moment different from many earlier AI breakthroughs.
The technology is no longer confined to generating information about biology. It is beginning to participate in the design of biology itself.
