Why AI-enabled bioweapons are raising alarm among researchers

A report from Anthropic has acknowledged that people have used its artificial-intelligence models to investigate ways of making chikungunya more transmissible, increasing the danger posed by bird flu and mapping venom-related toxins. The disclosure has renewed concern that AI systems could help people develop biological weapons, even as scientists differ sharply over how capable the technology is today.

MIT Technology Review reports that warnings from within the AI industry have become increasingly direct. Anthropic chief executive Dario Amodei recently said the technology carries serious risks and argued that development should slow. OpenAI chief executive Sam Altman responded that frontier AI should be advanced at a measured pace.

The debate intensified after AI researcher Jacob Coxon announced his departure from Anthropic. Coxon, who had also worked at OpenAI, said neither company was behaving responsibly in the face of the possible consequences. Anthropic employee Evan Hubinger publicly supported that view and said he personally considered the probability of AI killing all humans within the next decade to be greater than 10 percent.

How AI could lower the barriers

The concern is not limited to autonomous machines or science-fiction scenarios. Researchers worry that AI could assist with the discovery, design or release of biological and chemical agents. Potential threats could include a virus that harms people with particular genetic traits, a fungus capable of destroying crops, or an undetectable poison introduced into a water supply.

A 2022 experiment illustrated one route to such misuse. Scientists at Collaborations Pharmaceuticals had built an AI system to search for molecules that might become medicines. When they altered its purpose, the system produced 40,000 molecules in under six hours that could potentially be used as chemical warfare agents. Some were predicted to be more toxic than known nerve agents.

That result alarmed David Magnus, a Stanford professor who has studied the misuse of medical research and biotechnology since the late 1990s. He told MIT Technology Review that the experiment was deeply troubling, particularly as AI capabilities have expanded since then.

Modern language models can answer questions across a wide range of scientific fields. Dunja Sabra, a biosecurity researcher at the University of Hamburg, said these systems draw on an enormous body of accumulated scientific knowledge. They can also provide experimental guidance and instructional videos, potentially helping users move from general information to practical laboratory work.

At the same time, gene editing and synthetic biology have become more accessible. The growth of do-it-yourself biology has allowed some people to establish laboratories outside traditional research institutions. Sabra’s concern is that a determined individual could eventually combine AI assistance with these tools and succeed in creating something dangerous.

Existing protections have limits

Several safeguards are already used to reduce the risk. Companies that supply DNA generally screen orders for sequences associated with dangerous organisms. Researchers may also subject proposed work to adversarial reviews, asking independent experts to identify ways it could be misused and then considering measures to reduce those risks. AI developers have modified their systems to block requests for information that could enable harmful experiments.

None of these measures is guaranteed to work. Magnus described an ongoing contest in which defenders improve monitoring and screening while malicious users look for ways around them. He believes AI may eventually be needed to help detect and restrict harmful uses of AI itself.

Kevin Esvelt, an MIT biologist known for work on gene-drive technology and methods for containing it, added to the concern in a post on X. He said a large language model had revealed a type of bioweapon whose possibility he had not previously recognized, and urged people to adopt a cautious approach.

The Anthropic report shows that attempts to explore dangerous biological capabilities are not merely hypothetical. It does not, however, establish that current AI systems can independently design and deploy a successful pathogen. That distinction is central to the disagreement among experts.

Scientists disagree over the scale of the threat

At a media briefing, biologists at Imperial College London argued that existing AI tools remain too limited to complete the development of a bioweapon. Creating and testing a novel pathogen still requires difficult, lengthy work by people in laboratories, they said. Some researchers believe current safeguards are adequate for the risks that can realistically be managed today.

Wendy Barclay, an Imperial infectious-disease professor, also pointed to a more immediate source of pandemic danger: pathogens already circulating in the natural world. H5N1 bird flu has killed millions of birds and spread among dairy cattle in the United States. The virus was also found in captive mink at a Utah farm last month.

Sabra takes a longer view, looking five to 10 years ahead rather than only at present capabilities. She argues that governments should reinforce health systems, prepare treatments for known toxins and maintain medicine stockpiles. Her message is simple: “We need to be prepared.”