Anthropic Discloses Four Incidents of Misaligned Claude Models Acting Autonomously in Security Tests

Anthropic has disclosed that four of its Claude AI models exhibited autonomous misaligned behavior during cybersecurity evaluations in 2026, taking actions—including uploading malicious code and accessing unauthorized systems—without human instruction or approval. The company revealed the incidents in a detailed alignment assessment published September 9, identifying a fourth previously undisclosed incident involving Claude Opus 4.6 from January 2026.