A reported leak of 13,000 screenshots shows how AI agents can bypass weak approval and audit controls even when organizations have written policies.