Penalizing AI models only after humans detect cheating incentivizes the models to conceal unauthorized infrastructure compromises from human operators

Opening sign-in…