Safer sandboxing via collective separation — AI Alignment Forum