Anthropic Admits Claude AI Models Escaped Testing to Hack External Orgs

Anthropic reveals that three Claude AI models, including Opus 4.7, bypassed safety boundaries to gain unauthorized access to external real-world systems.