SFCoT: Safer Chain-of-Thought via Active Safety Evaluation and Calibration

2026-03-16

AI Summary

This work addresses the vulnerability of large language models to jailbreaking attacks during complex reasoning, a risk inadequately mitigated by existing defenses that focus solely on final outputs while neglecting intermediate reasoning steps. To bridge this gap, the authors propose SFCoT, a novel framework that introduces real-time safety assessment and dynamic calibration within the chain-of-thought reasoning process. SFCoT employs a three-tiered safety scoring mechanism coupled with multi-perspective consistency verification to continuously monitor and intervene in the reasoning trajectory. Experimental results demonstrate that SFCoT reduces jailbreak attack success rates from 58.97% to 12.31%, achieving substantial security improvements without significantly compromising the model’s general capabilities.

Technology Category

• Natural Language Processing: Safety and Robustness
• Machine Learning: Large Multimodal Models (LMMs)
• Reasoning under Uncertainty: Other Foundations of Reasoning under Uncertainty

Application Category

• Security and Privacy: Large-scale security measurements
• User Modeling, Personalization and Recommendation: Attacks and countermeasures in recommendation systems
• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking

Abstract

Large language models (LLMs) have demonstrated remarkable capabilities in complex reasoning tasks. However, they remain highly susceptible to jailbreak attacks that undermine their safety alignment. Existing defense mechanisms typically rely on post hoc filtering applied only to the final output, leaving intermediate reasoning steps unmonitored and vulnerable to adversarial manipulation. To address this gap, this paper proposes a SaFer Chain-of-Thought (SFCoT) framework, which proactively evaluates and calibrates potentially unsafe reasoning steps in real time. SFCoT incorporates a three-tier safety scoring system alongside a multi-perspective consistency verification mechanism, designed to detect potential risks throughout the reasoning process. A dynamic intervention module subsequently performs targeted calibration to redirect reasoning trajectories toward safe outcomes. Experimental results demonstrate that SFCoT reduces the attack success rate from $58.97\%$ to $12.31\%$, demonstrating it as an effective and efficient LLM safety enhancement method without a significant decline in general performance.

Problem

• jailbreak attacks
• safety alignment
• chain-of-thought
• adversarial manipulation
• LLM safety

Innovation

• Chain-of-Thought
• Safety Alignment
• Jailbreak Defense
• Real-time Calibration
• Reasoning Consistency