Meta let an AI agent handle sensitive user data, leading to a major internal security event. The error came as a result of an engineer implementing code on the bad advice given to them by the AI agent. How is Meta defending itself? "a human could also give erroneous advice."