César de la Fuente is a presidential associate professor and director of the Machine Biology Group at the University of Pennsylvania.
In my laboratory at the University of Pennsylvania, we use artificial intelligence to search for new medical countermeasures against infectious threats. The promise of the technology we use is revolutionary, offering a new range of potential antibiotics to combat the growing threat of resistant bacteria. But I am increasingly concerned that advances in AI could make it easier to create the very dangers we are trying to protect people from.
An AI assistant can do something a scientific paper cannot: respond to a reader’s confusion. It can explain unfamiliar concepts, connect findings across disciplines and help interpret why an experiment failed. That assistance could open scientific work to people who lack access to expert mentors. But it could also help someone with harmful intentions overcome gaps in knowledge that once might have stopped them from acting on those intentions.
The danger that worries me is the growing gap between access to biological expertise and our ability to protect people from its misuse. A digital tool can become available to millions of people almost immediately. The laboratories and public health systems needed to protect those people take years to build. We need a national biodefense strategy that takes this difference in speed seriously.
The time to act is now. Earlier this month, the World Health Organization sought information about the unexplained death of a worker at a plague research institute in Russia. Although the cause of death remains unclear, the situation highlights broader questions of preparedness and transparency.
There is evidence that some barriers to misuse of AI are eroding. In 2025, Britain’s AI Security Institute reported that the odds that nonexperts using advanced models could write a feasible plan for a virology experiment are roughly five times higher than for those using the internet alone. This was a test of experimental planning, not proof that an amateur could produce a biological weapon. But it suggests that AI can offer assistance beyond an ordinary online search.
The 2026 International AI Safety Report describes a related warning: Several major AI developers introduced additional safeguards after they could not rule out the possibility that their systems might meaningfully help novices develop biological weapons.
Producing a biological weapon still requires physical materials and practical capabilities that a convincing answer on a computer screen does not demonstrate. But uncertainty about how much harm AI could enable is a poor reason to postpone preparations that will take years.
The United States has policies to build on. The National Biodefense Strategy updated in 2022 addresses deliberate, accidental and naturally occurring threats. The White House’s 2025 AI Action Plan calls for stronger screening of synthetic genetic material and evaluations of national security risks. These efforts need to be brought together in a funded program that can test whether the country is prepared for biological misuse of AI, and fix whatever weaknesses it finds.
The White House should update the national strategy with that goal, identifying a senior official who would be accountable for its execution and would set public milestones. Technology companies and universities need to work with public health and national security agencies on a common assessment of the risks. Congress should fund the work over multiple years and require an annual accounting of progress, including unresolved vulnerabilities and the agencies responsible for fixing them.
AI developers should give qualified independent evaluators secure access to test models with potentially dangerous biological capabilities before release and after significant updates. Evaluations must measure how much assistance a system provides beyond existing tools and whether its safeguards withstand attempts at misuse. Developers should report serious safety failures to designated authorities. The public deserves a meaningful account of the findings, although details that could enable harm should be withheld.
Universities should accept independent scrutiny of the biological tools we develop. As a scientist, I have a strong stake in keeping research open: Researchers need to be able to reproduce one another’s findings. Any restrictions should rest on evidence of a specific risk and should remain subject to review. Government support for smaller laboratories would help prevent security requirements from making responsible research a privilege of wealthy institutions.
Even effective safeguards cannot substitute for the ability to respond when something goes wrong. Regular national exercises should test the response to AI-assisted biological threats, from the first unusual infection in a hospital through the delivery of effective protection. Researchers, hospitals, manufacturers and local health departments should rehearse together, testing the handoffs between them. The exercises should measure how long each step takes and identify who is responsible for correcting delays.
Congress must sustain this capacity between crises. Laboratories need staff and equipment to evaluate candidate medicines when demand arises. Public health departments need people who can recognize and investigate unusual patterns of illness. Such investments would serve us even if an AI-assisted biological attack never occurs. They would strengthen our response to resistant infections and naturally occurring outbreaks while giving the country regular opportunities to test its preparedness.
The public also belongs in these decisions. Patient groups and community representatives should have a formal role in deciding how health data are used and how scarce treatments would be allocated. They should also be able to see what preparedness exercises reveal and whether officials correct the failures.
The U.S. should pursue the same cooperation abroad, through common safeguards and agreements for sharing warning signals and distributing treatments. Helping other countries prepare is part of protecting Americans.
We still have time to make deliberate choices about how to protect people against the misuse of AI. We should not spend it waiting to learn how much easier AI has made it to harm them.
The post AI capabilities mean America needs a Manhattan Project for biodefense appeared first on Washington Post.
