Anthropic caught scientists asking Claude for help with bioweapons research

Anthropic, the company behind Claude AI, recently flagged and blocked multiple scientists attempting to use its AI models for potential biological weapons development. A new threat report released by the company details extensive misuse of its Claude AI platform. The investigation highlights five specific case studies involving working scientists exploring highly dangerous pathogens. Detecting this type of misuse remains highly complex, as legitimate biomedical research often closely mirrors the development of a bioweapon.

In one notable case from May, Anthropic’s internal biological safety classifier intercepted a request to draft a grant application for gain-of-function research on the chikungunya virus. This mosquito-borne pathogen causes debilitating symptoms and currently has no licensed treatment. The proposed research sought to increase the virus’s transmissibility and its ability to evade immune responses.

The affiliation of this research with a military institute immediately escalated the company's concerns. Anthropic also stopped a separate attempt involving gain-of-function research on bird flu, which focused on adapting the pathogen to spread more easily between mammals. Another intercepted query involved a researcher using Claude to map venom toxin peptides and optimize their characteristics.

Anthropic responded by permanently banning the accounts associated with these activities. The company utilized the findings to strengthen its safety guardrails against future misuse. The report emphasizes that newer models demonstrate increasing proficiency in handling sophisticated scientific tasks.

This capability makes them highly attractive targets for actors seeking a true research assistant in complex biological engineering. Anthropic chose to keep the identities and institutions of the involved scientists anonymous. The company stated that identifying these working professionals could expose them to physical harm, noting that the researchers may not have explicitly intended to cause harm.