Anthropic says it blocked 5 attempts to use AI for potential bioweapons development

Artificial intelligence company Anthropic said Thursday that it blocked five attempts to use its AI models for activities that could potentially contribute to the development of biological weapons.

“While evaluations are useful because they provide evidence of capability, they cannot concretely demonstrate that such capability would ever be used to develop biological weapons in the real world,” the company said in a report on the misuse of its AI systems.

The cases involved research related to chikungunya, avian influenza and smallpox, according to the report.

In one case, a researcher affiliated with an unidentified military institution used Anthropic’s Claude chatbot while preparing a research proposal involving chikungunya, a mosquito-borne virus.

Although the research could have legitimate purposes, including vaccine development, the researcher also sought assistance that raised concerns about creating more dangerous variants of the virus.

Other cases involved efforts to use AI to explore ways of increasing the transmissibility of avian influenza and to gather information about animal toxins that could potentially be used in weapons.

The company also found that some users tried to bypass its safeguards to access its models, using fraudulent accounts, private networks and third-party resellers.

Beyond biological research, the report identified other potentially malicious applications of AI, including the development of conventional weapons and cyberattacks.

One case involved queries from Yemen concerning the development of a guided rocket using mobile technology, according to the report.

At the same time, Anthropic cautioned that identifying potentially dangerous capabilities does not establish that users intended to develop biological weapons or would have succeeded in doing so.

The findings come amid growing concerns over the ability of increasingly capable AI systems to facilitate harmful activities that previously required specialized expertise.

Anthropic has introduced stronger safeguards for sensitive areas such as biology and cybersecurity, while access to some of its most capable biology-focused models is restricted to vetted organizations.