In his annual speech on threats to the nation, Sir Ken McCallum, the head of the UK’s intelligence service, MI5, warned that AI had the potential to “change the stakes” on existing and even new bioweapons. Here we look at how AI might help in the development of bioweapons and the hurdles that remain.
In one case, Anthropic said researchers used the chatbot to plan experiments that would introduce mutations in a highly pathogenic bird flu virus to make it adapt and spread more easily in mammals. The work, and the other cases, could signify attempts to make pathogens more dangerous, but such studies are also done in legitimate labs to help researchers spot future pandemics and develop vaccines and drugs to treat disease.
A flurry of studies has shown how AI tools can provide potentially dangerous information on bioweapons. In a recent exercise at MIT, non-science students were asked if they could get an AI chatbot to help them unleash a pandemic. Within an hour, the chatbots listed four potential pandemic pathogens, explained how to make them from synthetic DNA, and supplied names of DNA synthesis companies that were unlikely to screen their orders. It is easier said than done, however.
In another case, the Stanford University researcher Dr David Relman described how an AI explained how to modify a pathogen to make it resistant to treatment. Another scientist revealed that an AI chatbot had explained how to use a weather balloon to release biological weapons from the sky.
The dual-use nature of biological research can make it hard to distinguish valuable work that could protect health from attempts to develop bioweapons. In 2025, scientists at Los Alamos National Laboratory trained an AI to predict how exceptionally potent neurotoxins from marine snails latch on to human cells, work that could help in the development of antidotes. But the US government became concerned when Chinese researchers drew on the work to develop their own tool to design new toxins. The Chinese researchers said the work was aimed solely at developing drugs.
In April, scientists raised further concerns when they made the first viruses designed by AI. The viruses are a particular type that only invade bacteria, and the work could lead to new medicines, but experts warned that it raised “urgent biosafety and biosecurity questions”.
Most work on bioweapons is beyond the reach of unskilled groups: as well as being highly technical, it is extremely dangerous. While AI chatbots might be prompted to give information on bioweapons, a list of instructions is not enough. “Information is only one part of the problem,” says Lentzos. “Tacit expertise, access to biological materials and equipment, experimental validation, scale-up and delivery all remain substantial barriers.”
Sonia Ben Ouagrham-Gormley, the deputy director of the biodefense programme at George Mason University in Virginia says tacit knowledge – the unwritten knowhow that an AI is unlikely to help with – has been key in the success of historical state-run bioweapons programmes.
Prof Martin Pačesa at the University of Zurich said access to material was critical. “Any graduate physicist could sketch up plans for a potentially more effective nuclear bomb, but where is he going to get a lump of uranium from? Same for bioweapons. If you have access to material and facilities capable of producing them, then AI is not going to make too much difference if you want to do harm.”
Some AI companies are taking steps to prevent people from using their tools to develop bioweapons, but the majority are not announcing what, if anything, they are doing, says Prof Tom Inglesby, director of the Johns Hopkins Center for Health Security. “We have no formal government policies in place in any country yet to protect against this kind of misuse,” he says.
He wants formal independent evaluations of new AI models before they are released, with any dangers eliminated before the models are made public. And high-risk biological data, such as information on how pathogens might spread more easily, should be restricted. Meanwhile, companies that make genetic material for researchers must screen orders to ensure they are not for the creation of novel, highly dangerous viruses. Some screening already happens, but it is largely voluntary, and can be sidestepped if researchers use AI to design new, harmful genetic sequences that companies are not checking for.