Anthropic has flagged and stopped multiple scientists using its AI models to create potential biological weapons, according to a new report shared on Thursday. The report details five case studies where Claude was misused for biological weapon research, including gain-of-function studies on chikungunya virus and bird flu, and development of venom toxin peptide atlases. Anthropic's biological safety classifier detected these requests, prompting account bans and model safeguard improvements.

The company emphasized the difficulty in distinguishing legitimate vaccine research from harmful applications, noting that intent is often unclear. Jacob Klein, Anthropic's head of threat intelligence, described the situations as 'incredibly nuanced' and stressed that the company erred on the side of caution due to potential consequences. Anthropic did not name individuals or institutions involved, citing concerns for their safety and the lack of asserted harmful intent.

The report also covers misuse in surveillance, software exploits, propaganda, and weapons systems. Findings are being used to enhance model protections against future abuse.