AI Agents Can Get Approval and Still Do the Wrong Thing

Human approval is not a security boundary unless it is bound to the exact action being approved,...