The experiment was simple in design but extraordinary in execution. Researchers at Stanford University and the Arc Institute trained two genome language models, Evo 1 and Evo 2, on roughly 9 trillion DNA base pairs from plants, animals, microorganisms, and viruses. They then asked the models to design genomes for a bacteriophage — a virus that infects bacteria — based on the well-studied Phi X-174, which only attacks E. coli and is harmless to humans.
The AI generated about 700,000 candidate genomes. The researchers selected 285, synthesized them as DNA, and inserted them into E. coli. Sixteen produced working viruses that replicated, destroyed bacteria, and spread to infect other cells. Some replicated even faster than the natural Phi X-174.
The result that matters: a mixture of the AI-designed phages killed E. coli strains that had become resistant to naturally occurring phages. This is not just a laboratory curiosity. This is AI solving a problem that has defied conventional approaches for decades.
Bacteriophages have been proposed as a solution to antibiotic-resistant bacteria for roughly a century. The problem has always been matching the right phage to the right infection — a process that requires screening and testing that can take weeks, while the patient runs out of time. AI that can design phages on demand changes that calculus entirely.
The researchers' findings suggest a path forward: AI-guided generation could eventually enable "more durable phage-based therapies" that can adapt as bacteria evolve resistance. Instead of searching for a natural phage that happens to work, doctors could call up an AI model and ask it to design a phage optimized for the specific bacterial strain in front of them.
But the media narrative has been dominated by fear. The headlines: "AI used to create new viruses" — "urgent biosafety warning" — "it could be the beginning of a dangerous new trend."
The safety concerns are real. Johns Hopkins experts Thomas Inglesby and Moritz Hanke, writing in an accompanying commentary in Science, put it bluntly: "The ability to compose viral genomes using generative AI now exists; the governance to safely steer it does not." They warned that AI-designed genomes "might encode new pathogens that cannot be contained by existing countermeasures."
The Stanford team took precautions. They excluded viruses that infect humans from the training data. They also removed viruses from plants, animals, and fungi that possess human infection–like viruses. They used a phage that only attacks E. coli. But as experts point out, there is no guarantee that every other scientist attempting similar work will be so careful. And the regulatory frameworks have not kept pace. AI-based pathogen design on computers may fall outside existing rules as long as it does not target specific hazardous agents.
There is a more nuanced problem that cuts both ways. The researchers' work is important precisely because it demonstrates that AI can design functional genomes — but it also shows how much the technology relies on high-quality training data and experimental validation. The threat from "full AI design and writing of a genome of a virus or bacteria" has been described by some experts as "very overblown" compared to the much easier path of making gain-of-function changes to existing pathogens.
The real risk isn't that an AI model will suddenly invent a supervirus. It's that the capability is now out there — and the safeguards that prevent its misuse are voluntary, not mandatory. The Stanford team's own paper urges future researchers to "consult both safety and security professionals throughout the project lifecycle." But that's a recommendation, not a rule.
The irony is that the same concerns about biosecurity are now being used to justify restrictions that could slow down the very medical research that phage therapy desperately needs. Antibiotic resistance is projected to cause 10 million deaths annually by 2050. AI-designed phages might be one of the few tools that can stop that trajectory. But if the regulatory response is to lock down access to the models and the data, the patients who need those treatments will pay the price.
P.S. The Oxford researcher who commented on this study put it simply: "If AI can directly design viruses optimized for specific functions, it could become a useful biotechnology tool." The technology isn't good or bad — it's just capable. And that's the thing we're still learning to handle.
Frequently Asked Questions
Q: What exactly did the Stanford researchers create with AI?
A: The researchers used genome language models Evo 1 and Evo 2 — trained on roughly 9 trillion DNA base pairs across all domains of life — to generate thousands of new viral genomes from scratch. They selected nearly 300 designs, synthesized them as DNA in the lab, and found that 16 produced fully functional viruses that could infect and kill bacteria.
Q: Can these AI-designed viruses infect humans?
A: No. The viruses are bacteriophages — viruses that only infect bacteria and are completely harmless to humans, animals, and plants. The Stanford team deliberately excluded viruses that infect humans, animals, plants, or fungi from the AI's training data.
Q: How did the AI-designed viruses perform against drug-resistant bacteria?
A: A mixture of the AI-generated phages successfully killed E. coli strains that had evolved resistance to naturally occurring phages. A comparable mixture of natural phages could not. The researchers say this "lays out a path for generating adaptive and resilient phage therapies against rapidly evolving pathogens".
Q: What are the biosecurity concerns around this research?
A: Johns Hopkins experts Thomas Inglesby and Moritz Hanke, writing in an accompanying Science commentary, put it bluntly: "The ability to compose viral genomes using generative AI now exists; the governance to safely steer it does not". They warned that AI-designed genomes "might encode new pathogens that cannot be contained by existing countermeasures". The researchers took precautions — but these are voluntary, not mandatory.
Q: Why aren't existing regulations enough to prevent misuse?
A: Current rules typically require screening DNA synthesis orders against known pathogens. But AI can generate novel sequences that are functionally dangerous yet don't match any database entry, creating a "zero-day" problem in biosecurity. Additionally, AI-based pathogen design done entirely on a computer may fall outside existing regulations as long as it doesn't target specific hazardous agents.
Q: Is the risk of AI creating a deadly virus imminent?
A: Experts say that's a major leap from what has been demonstrated. Imperial College London professor Tom Ellis noted the phage used in the study has a very small genome (about 5,400 DNA letters) — the simplest possible case. For comparison, SARS-CoV-2's genome is six times longer. Barcelona's Jordi García Ojalvo added that the efficiency rate is low (just 16 viable from hundreds of thousands of designs), making it "difficult to imagine these models automatically generating viable genomes 'out-of-the-box'".
Q: Is Evo 2 available for anyone to download?
A: Yes. Evo 2 is open source, with model weights and code available on GitHub and HuggingFace. There are several versions available, including 1B, 7B, 20B, and 40B parameter models, with the 7B version capable of running on any supported GPU without requiring Transformer Engine.
Q: Who funded this research?
A: The work was conducted by researchers at Stanford University and the Arc Institute, a nonprofit research organization. The study was published in the journal Science on August 6, 2026.
