Competitive safety via gradated curricula — AI Alignment Forum