Anthropic reveals fourth security breach involving Claude model

Anthropic has disclosed a fourth cybersecurity incident involving its Claude model, which occurred in January 2026. The breach was identified during a review of logs shared with the Model Evaluation and Threat Research Organization (METR) in August. The incident involved an early version of Claude Opus 4.6, which was mistakenly connected to the open internet during a cybersecurity test. Anthropic expanded its investigation to include approximately 481 million logs, reconfirming the four incidents without finding additional severe cases. The breaches occurred within the same cybersecurity assessment process conducted by an external testing organization, where the model was incorrectly linked to the internet due to a configuration error. The company emphasized that standard security measures were not applied during the testing phase. The initial incident, reported on July 30th, involved about 141,000 session logs. However, automated screening missed some logs connected to the external network. Anthropic's expanded review, which included a second deep analysis using Claude on 9.2 million high-risk logs, confirmed the four incidents. The company stated that no other cases of similar or more severe severity were found.