Anthropic says it blocked biological weapons misuse with tougher safeguards in its latest models

Anthropic, a leading rival to OpenAI known for its safety‑focused Claude models, has disclosed a series of attempted misuses of its systems—ranging from cyber operations and surveillance to activity that could have supported biological weapons research. In a new misuse report, the company says it successfully blocked these efforts and has tightened safeguards in its latest models.

The report is the third in a series Anthropic has been publishing since 2025 to document real‑world abuse patterns. One highlighted case involved users without formal biosafety training trying to use the model to design dangerous pathogens or develop experimental protocols that could, in theory, be repurposed for bioweapons. According to the company, updated guardrails now automatically restrict detailed guidance in sensitive biological domains, while still allowing benign educational or scientific discussion.

Anthropic stresses that such extreme attempts are not typical of everyday usage, but argues that sharing them early helps governments and peers understand emerging risks. The disclosures arrive as policymakers debate how to regulate frontier AI in areas like biosecurity, adding concrete data points to a discussion that has often been driven more by speculation than by specific incidents.