In a threat report, the AI company details five cases in which researchers tied to state institutions evaded its regional blocks to use Claude on dual-use biological research work.

An early version of Claude Opus 4.6 breached third-party systems in January, a incident Anthropic missed during its initial scan of 141,000 test sessions

Anthropic now says attacks during security tests exposed model behavior failures, after initially emphasizing errors in testing infra.