As artificial intelligence systems grow more capable of analyzing genetic code, more than 100 researchers from leading universities are warning that some publicly available biological data could be used to design dangerous viruses.

Researchers from Oxford, Johns Hopkins, Fordham, Columbia, New York University, Stanford, and the University of Michigan published a paper Feb. 6 in the journal Science endorsing a framework to govern certain biological data in a manner similar to how sensitive health records are handled.

The warning comes as AI models are increasingly being used in drug discovery and biological research, with some systems trained to interpret DNA sequences much like large language models interpret human language.

The authors argued that although today's consumer chatbots are not capable of designing pathogens, future models trained on sensitive virology data could lower the barrier for misuse if governments do not establish clear data-access rules.

"Fortunately, governments can reduce the biosecurity risk posed by AI models by improving on a long-standing policy from the domain of privacy protection: data-access rules," the researchers wrote.

"Scientists broadly accept data-access limitations to protect privacy. Government programs in the United States, Europe, and elsewhere mediate access to personally identifiable and sensitive data to permit rigorous research without compromising privacy."

"Although imperfect, these frameworks have shown that responsible governance and scientific progress are not contradictions. Governments can apply a similar approach to improve biosecurity today."

The authors are not calling for broad restraints on science and argue that most biological data should remain open.

Instead, they wrote that only a narrow subset of pathogen-related data that could materially increase the risk of misuse should be restricted.

"By limiting public access to a narrow band of new and sensitive pathogen data, governments may reduce the risk of future AI model misuse without substantially hampering scientific progress," they wrote.

"Governments should also ensure that biosecurity concerns do not exclude researchers from less wealthy institutions or nations from contributing to or benefiting from scientific advances."

Jassi Pannu, assistant professor at the Johns Hopkins Center for Health Security and one of the report's authors, told Axios the concern is not about off-the-shelf AI systems such as OpenAI's ChatGPT or Anthropic's Claude.

Some AI models designed for biological research use architectures similar to large language models but are trained on genetic sequences rather than text.

Pannu noted that some developers have voluntarily chosen not to train their models on virology data out of concern about making such capabilities widely available. However, if sensitive data remains publicly accessible, third parties could fine-tune models on that material without following similar safeguards.

"Legitimate researchers should have access," Pannu said. "But we shouldn't be posting it anonymously on the internet where no one can track who downloads it."

She added that there is no expert-backed framework clearly identifying which biological datasets pose meaningful risks, leaving some frontier AI developers to make voluntary decisions about excluding viral data from training.

The report also warns that new biological AI models are sometimes released "without conducting basic safety assessments" that would be standard in other high-risk scientific fields. It calls on governments to regularly reassess any restrictions and refine them as AI capabilities evolve.

"It's been shown time and time again that we don't do a good job of predicting AI capability trends," Pannu said. "We're constantly surprised."

"And so, I would argue that for these large-scale, consequential risks, we should try and prevent these worst-case scenarios and be prepared for them."

"It's not necessarily that I'm saying that I think this will happen and I know exactly when it will happen, but I think it's worth trying to prevent the worst-case scenario, even if we're unsure exactly when it might happen," added Pannu.

• Congress Targets Sales of Possibly Risky Synthetic DNA
• Trump to UN: Will Use AI to Enforce Bioweapons Treaty