Anthropic just published a case study where threat actors jailbroke Claude and used it to run entire attack campaigns autonomously.