Just days ago, a study published in the journal Science sent shockwaves through the scientific community. 

Researchers from Stanford University and the Arc Institute in California used generative AI to design entirely new, functional viruses from scratch. Of the hundreds of designs the AI generated, 16 proved viable—capable of infecting and replicating inside bacteria.

It is a landmark achievement. It is also the kind of headline that keeps biosecurity experts awake at night.

How the AI Did It

The researchers used two genome language models, Evo 1 and Evo 2, trained on roughly nine trillion nucleotides from millions of viruses, bacteria, plants, and animals. 

These models work much like large language models, like ChatGPT. But instead of predicting words, they predict genetic sequences.

The team asked the AI to generate complete genomes for a bacteriophage, a virus that infects and destroys bacteria. The AI churned out roughly 700,000 potential designs. The researchers selected 285 promising ones, synthesised the DNA, and inserted them into E. coli bacteria.

Within hours, 16 petri dishes began showing clear spots—the unmistakable sign that new viruses had emerged and were actively killing their bacterial hosts. 

Some of the AI-designed phages even outperformed the natural version, killing antibiotic-resistant strains of E. coli that the original could not.

"We didn't add anything," said Brian Hie, an assistant professor at Stanford and co-author of the study. "The model generated the entire genome end-to-end in a single left-to-right pass."

A New Era for Medicine

More than 2.8 million antimicrobial-resistant infections occur in the United States each year, killing over 35,000 people annually. 

AI can change that. The ability to rapidly design and tune phages to overcome bacterial resistance could "transform phage therapy," the researchers wrote. A cocktail of multiple AI-generated phages could make it far harder for bacteria to develop resistance.

The researcher imagines a future where AI-designed viruses fight multi-drug-resistant infections, treat cancer, or even help with Alzheimer's. 

The lab spontaneously broke into applause when the results came in. It was, the researcher said, the first time that had ever happened.

Pandora's Box Is Now Digital

But the same technology that can design life-saving therapies can also design biological weapons.

The viruses created in this study pose no threat to humans. They are bacteriophages that infect only bacteria. The researchers deliberately excluded viruses capable of infecting humans, animals, or plants from the AI's training data. All experiments were conducted in secure laboratories.

But these safeguards are voluntary. As Johns Hopkins University public health researchers Thomas Inglesby and Moritz Hanke wrote in a commentary accompanying the study: "The ability to compose viral genomes using generative AI now exists; the governance to safely steer it does not".

They warned that the models' safety measures "can be partly circumvented by fine-tuning the models on pathogen data". Future AI agents might even coach non-experts on biological design, enabling a wide range of bad actors to create novel bioweapons.

Kevin Esvelt, a genetic engineer at MIT who studies biological threats, put it bluntly: "Better is just to prepare for the world where it's easy to cause pandemics".

The Governance Gap

The research was first published as a preprint in September 2025. The AI models are open-source and have already been fine-tuned for other tasks. The frontier is moving much faster than regulation.

More than 100 researchers have endorsed an entreaty to prevent AI systems from being used for "deliberately harmful applications such as bioweapons development". A bipartisan bill in the US Senate would require companies that make and sell custom DNA sequences to screen orders for dangerous sequences.

Biosecurity expert Filippa Lentzos from King's College London argues that regulation should focus on the point where digital genetic instructions become physical DNA. "The challenge," she said, "is to connect that existing governance to the new upstream capability to design biology digitally".

New Era of Biomedical Field

AI has crossed a threshold. It can now design functional, replicating viruses from scratch. The same capability that could cure antibiotic-resistant infections could also, in the wrong hands, create the next pandemic.

The researchers took responsible precautions. But as Hie himself acknowledged: "If you think about it, nature itself is constantly producing viruses with pandemic potential". The difference is that nature does not have intent. Humans do.

The question is no longer whether generative viral genome design will exist. It already does. The question is whether we can build the governance to ensure it is used for good—before someone decides to use it for something else.

Because once the blueprint is out, you cannot un-invent it. And the lab that spontaneously burst into applause may, one day, be remembered not for the lives it saved, but for the door it opened.