Chain of Thought Monitorability: A New and Fragile Opportunity for AI Safety — AI Alignment Forum