Reasoning Models Struggle to Control Their Chains of Thought

OpenAI finds reasoning models can't easily hide their thinking — good news for safety monitoring but reveals limits of CoT steering.