How nice that state-of-the-art LLMs reveal their reasoning ... for miscreants to exploit

Researchers introduce H-CoT, a prompt-based jailbreak that hijacks chain-of-thought safety reasoning in leading LRMs (OpenAI o1/o3 and o3-mini, DeepSeek-R1, Gemini 2.0 Flash), using a Malicious-Educator dataset to drastically lower safety rejection rates and expose delayed/weak safeguards; tests conducted via web UIs/APIs underscore differences between remote filtering and local deployments, while external vendor evaluations of DeepSeek vary in rigor—collectively suggesting exposed CoT reasoning is a significant new safety attack vector.