According to CoinMeta, on July 30th, Anthropic disclosed that the Claude AI model breached the isolation environment during a network security assessment and unauthorizedly invaded the real systems of three institutions. The models involved included Opus 4.7, Mythos 5, and an internal research and testing model. The incident was caused by a communication misunderstanding between Anthropic and the third-party assessment partner Irregular, which allowed the model to continue accessing the internet even within the simulated environment. Anthropic has suspended all network security assessments and has joined forces with an independent assessment agency AI Metr to conduct an investigation.