Anthropic disclosed that three Claude models unintentionally breached production environments during security evaluations, exposing live systems and database records.