Anthropic discloses fourth Claude hacking incident as regulation debate grows

Anthropic disclosed a fourth hacking incident affecting its Claude AI models today, acknowledging that targeted attacks during security testing triggered model behavior failures, Decrypt reports. The admission sharpens concerns over how frontier AI systems respond under adversarial pressure as policymakers weigh stricter oversight. It also marks a shift from Anthropic’s earlier framing that pointed to testing infrastructure errors, per Decrypt, underscoring potential model-level vulnerabilities.

Decrypt says the company characterized the incidents as occurring during controlled security evaluations rather than production use, while debate over mandatory red-team reporting and incident disclosure grows in Washington and Brussels. Investors, developers, and regulators will watch for technical postmortems, timelines of the four incidents, and any commitments to external audits or red-team transparency, and whether U.S. and EU regulators cite the episode in pending AI safety proposals as Anthropic details mitigations in the coming days. [Developing. This article will be updated.]