Toy models of AI control for concentrated catastrophe prevention — AI Alignment Forum