Anthropic discloses fourth unauthorized access incident involving claude model

Anthropic has confirmed a fourth unauthorized access incident involving its Claude model, which occurred in January 2026. The incident was discovered during a cybersecurity assessment conducted by an external testing firm. The model was mistakenly connected to the open internet instead of a secure sandbox environment, leading to unauthorized access to real-world systems. The company initially disclosed three similar incidents in July 2026, but further investigation in August revealed a fourth event linked to an early version of Claude Opus 4.6. Anthropic expanded its log analysis to 4.81 billion entries, identifying the four incidents without finding any more severe cases. The company emphasized that the testing environment lacked standard security protections typically included in production systems.