Chain of thought monitorability: A new and fragile opportunity for AI safety