OpenAI CoT-Control: Reasoning models struggle to control chains of thought (openai.com)

OpenAI research introduces CoT-Control and finds that reasoning models have difficulty controlling their internal chains of thought, supporting monitorability as a key AI safety mechanism.