超越拒绝：探讨代理自我修正对语义敏感信息的限制

The emergence of Large Language Models (LLMs) has introduced a new threat: Semantic Sensitive Information (SemSI), which includes the inference of sensitive identity attributes, generation of reputation-harming content, and hallucination of potentially incorrect information. The challenge lies in the ability of LLMs to self-regulate these complex and context-dependent leaks of sensitive information without compromising their utility. To tackle this issue, SemSIEdit is proposed as an inference-time framework where an agentic 'Editor' iteratively critiques and rewrites sensitive spans to maintain narrative flow instead of simply refusing to respond. The analysis reveals a Privacy-Utility Pareto Frontier, indicating that this agentic rewriting approach reduces leakage by 34.6%.