Anthropic disclosed attempts to use Claude for work with potentially dangerous pathogens and toxins between December 2025 and August 2026.

Scope of incidents

According to the company, malicious requests were disguised as legitimate research projects and originated from a mix of individual scientists and entities linked to military institutions of unnamed countries. The queries targeted a range of biological agents and toxic compounds, including avian influenza, orthopoxviruses, chikungunya and various toxic substances.

Detection and mitigation

Anthropic’s security team identified requests aimed at enhancing harmful properties of pathogens and began manually reviewing flagged dialogues. The company reports that it subsequently strengthened its safety filters and increased human oversight of sensitive interactions.

Operational measures

The report states that the pattern of misuse involved iterative prompts intended to refine experimental protocols and increase virulence or toxicity. In response, Anthropic implemented additional gating on high-risk topics and adjusted monitoring thresholds to reduce automated exposure.

Attribution and disclosure

Anthropic did not name specific organizations or countries involved; the company limited its disclosure to describe the types of users and the thematic focus of the requests. The timeline and affected topics were provided by Anthropic as part of the company’s security summary.