Scientists Use AI to Design 16 New Viruses That Kill Bacteria, Raising Biosecurity Concerns

Scientists at Stanford University and the Arc Institute have used artificial intelligence to design entirely new viruses capable of infecting and destroying bacteria, a scientific first that researchers say could open new paths for fighting drug-resistant infections while simultaneously reigniting debate over how quickly biotechnology is outpacing the safeguards meant to govern it.

The research, published this week in the journal Science, marks the first time scientists have used AI not simply to tweak existing viral genomes but to generate functional, previously unseen viruses from scratch, built entirely from patterns the AI learned by studying genetic sequences drawn from millions of organisms across the natural world.

Moving Beyond Copying Nature

For years, scientists have been able to synthesize viruses in the lab, typically to develop and test antiviral drugs and vaccines or to better understand how these microorganisms behave. That earlier work, however, relied almost entirely on replicating known pathogens or their existing variants. The new study represents a meaningful departure from that approach, using AI to generate genuinely novel genetic blueprints rather than working from a template drawn directly from nature.

The research team worked with bacteriophages, viruses that infect only bacteria and are increasingly viewed as a promising alternative to antibiotics in the fight against drug-resistant infections. Because bacteriophages have relatively small, comparatively simple genomes, they are easier to synthesize and manipulate under controlled laboratory conditions than viruses capable of infecting humans or other complex organisms.

How the AI Models Were Built

The viruses were generated using Evo 1 and Evo 2, foundational AI models developed specifically for computational biology applications. Both systems were trained on genetic sequences, rather than text, drawn from across all domains of life, allowing them to learn complex evolutionary patterns, including how genes are typically organized, which sequences tend to be conserved across species, and the underlying biological constraints that allow an organism to remain functional. According to researchers, the models were further trained specifically on genetic sequences from thousands of viruses belonging to the same broad family as Phi X-174, a well-studied bacteriophage capable of infecting the bacterium E. coli, while deliberately excluding any viral sequences capable of infecting humans, animals, plants or fungi from that additional training step.

Using Phi X-174 purely as a reference point, rather than a template to be copied, the AI models generated hundreds of thousands of candidate viral genomes engineered to retain the biological architecture necessary to recognize and infect E. coli, while differing considerably from any naturally occurring bacteriophage at the level of their actual genetic sequence.

From Hundreds of Thousands of Candidates to 16 Working Viruses

Researchers narrowed that enormous pool of AI-generated candidates down to roughly 300 genomes considered most likely to be functional, based on factors including gene organization and the presence of regulatory elements informed by the biology of Phi X-174. Each of those 300 candidate genomes was then physically synthesized, molecule by molecule, in the laboratory and introduced into E. coli bacteria to test whether it could actually produce a working virus.

Of those 300 synthesized genomes, only 16 gave rise to fully functional bacteriophages. Each of the 16 successful viruses featured previously unpublished genetic sequences, different genes, new regulatory elements, and in some cases entirely different genome sizes compared with anything found in nature. The viruses' behavior also varied considerably from one another, with some infecting bacteria more quickly than others and differing in their overall ability to replicate.

Testing Against Drug-Resistant Bacteria

Beyond simply confirming that the AI-designed viruses could function, researchers also tested their ability to combat bacterial resistance directly, a key question given the viruses' potential as an alternative to antibiotics. In that experiment, scientists exposed strains of E. coli that had already developed resistance to Phi X-174 to a mixture of the newly AI-designed phages alongside a comparison mixture of natural phages similar to Phi X-174. The AI-generated viruses were able to rapidly overcome the bacteria's existing resistance and establish infection, a result the study's authors said demonstrates what they described as a path toward AI-generated phage therapies capable of targeting rapidly evolving bacterial pathogens.

A Milestone With Two Very Different Implications

Researchers involved in the work say the approach could eventually support the development of highly personalized bacteriophage treatments capable of evolving nearly as quickly as the bacterial pathogens they are designed to fight, a potentially significant advance given the growing global public health threat posed by antibiotic-resistant infections.

At the same time, the achievement has renewed concern among biosecurity experts about the pace at which AI-driven genome design is advancing relative to the regulatory frameworks meant to govern it. While the viruses created in this specific study were deliberately restricted to targeting bacteria rather than humans, the underlying capability, using AI to design an entirely novel, functional virus from scratch, raises broader questions about how such tools might eventually be misused.

Moritz Hanke, a researcher at the Johns Hopkins Center for Health Security, warned about that dual-use risk directly, noting that a similar genome language model could in principle be asked to design something considerably more dangerous, such as an influenza genome modified to be more infectious or lethal. Speaking separately to The New York Times, Hanke said there remains what he described as a huge disconnect between the speed at which the underlying science and technology are advancing and the development of regulatory frameworks capable of keeping pace with it.

A Debate That Predates This Study

Concerns over AI's potential role in biological weapons development are not new. A study published roughly three years ago by the Rand Corporation warned that even the AI systems available at that time had the capacity to help refine the planning and execution of attacks using biological weapons. Rand separately cautioned that the speed at which AI systems continue to evolve routinely outpaces governments' ability to develop and implement effective regulatory oversight, a warning that has only grown more pointed as genome-design capabilities like those demonstrated in this new study have continued to advance.

With the research now published and drawing significant attention across both the scientific and biosecurity communities, calls are likely to grow for closer collaboration between AI developers, biologists, policymakers and regulators to establish clearer safeguards around genome language models before the technology advances further. In the meantime, the study's authors have framed their work primarily as a proof of concept for AI-assisted phage therapy, even as the broader scientific community continues grappling with how to balance the technology's genuine medical promise against the biosecurity risks it simultaneously introduces.