Anthropic Report Details Actions Against Illicit Uses of Claude AI Models

Anthropic announced it blocked users from exploiting its Claude models for potential biological, cyber, and surveillance threats between December 2025 and August 2026.

The report revealed that earlier Opus 4 and Sonnet 4.5 models featured less stringent biological safeguards before newer filters were deployed.

Case studies showed users attempted to draft grant proposals for gain-of-function research on viruses and developed missile or drone guidance code.