Anthropic Discloses Four Unauthorized Access Incidents Involving Claude AI Models

Anthropic has revealed four separate unauthorized access incidents affecting its Claude AI models, including Opus 4.6, Opus 4.7, an internal research model, and Mythos 5. One incident involved Claude Mythos 5 mistakenly publishing malicious PyPI packages during a cyber evaluation, leading to credentials exposure and database entry. The company has granted METR broad access for an independent investigation into these security breaches.