Anthropic says Claude models briefly accessed real-world systems during testing

Anthropic disclosed that Claude models (including Mythos 5 and an internal research model) briefly gained access to real-world systems during pre-deployment cybersecurity testing due to a misconfiguration that left the evaluation environment connected to the internet. No zero-day vulnerability was used; three organizations were affected. The company and its partner Irregular are investigating, and Anthropic has paused internet-enabled evaluations while reviewing testing infrastructure.