Anthropic said Thursday that it blocked attempts to use its artificial intelligence systems for cyberattacks, surveillance and research that could have supported biological weapons. The company published excerpts of malicious code and prompts in its third threat report since March 2025, describing the cases as notable and novel rather than typical use.
Anthropic said increasingly capable models can lower the expertise required for elaborate cyber operations and can give individuals access to assistance once limited to larger groups. The company reported adding stronger safeguards to its latest models to restrict biological research that could be adapted to weapons development.
Anthropic urged governments and competing AI companies to share information and prevent similar misuse, presenting the findings as an industry-wide security concern. Associated Press noted that outside experts continue to call for public regulation rather than relying only on companies to identify and police misuse of their own products.
A provider can observe account behavior, prompts and outputs on its service, but it may not see activity transferred to local systems or other models. Publishing examples can improve defensive awareness while also requiring care not to provide a detailed playbook for new attackers. Biological and cyber risk assessments distinguish information access from a user’s ability to obtain materials, equipment and operational capability.
At the edition deadline, the underlying cases and claimed effectiveness of the new safeguards were reported by Anthropic and had not been independently reproduced in full. The next documented developments will be external technical review of the report and any government response and new cross-company threat-sharing arrangements and measured changes in successful abuse attempts.