SFCoT: Safer Chain-of-Thought Proactively Monitors Intermediate Reasoning Steps for Jailbreak Attacks

arXiv 2603.15397·medium signal

SFCoT (Safer Chain-of-Thought) addresses a critical gap where existing LLM safety defenses only filter final outputs, leaving intermediate reasoning steps unmonitored and exploitable by adversarial manipulation. The framework proactively evaluates each CoT reasoning step against safety criteria and calibrates the generation trajectory before harmful outputs can emerge. Particularly relevant for agentic systems where multi-step reasoning is the primary attack surface for jailbreaking.