Beyond Refusal: Probing the Limits of Agentic Self-Correction for Semantic Sensitive Information

Abstract

While defenses for structured PII are mature, Large Language Models (LLMs) pose a new threat: Semantic Sensitive Information (SemSI), where models infer sensitive identity attributes, generate reputation-harmful content, or hallucinate potentially wrong information. The capacity of LLMs to self-regulate these complex, context-dependent sensitive information leaks without destroying utility remains an open scientific question. To address this, we introduce MethodName, an inference-time framework where an agentic “Editor” iteratively critiques and rewrites sensitive spans to preserve narrative flow rather than simply refusing to answer. Our analysis reveals a Privacy–Utility Pareto Frontier, where this agentic rewriting reduces leakage by 34.6% across all three SemSI categories while incurring a marginal utility loss of 9.8%. We also uncover a Scale-Dependent Safety Divergence: large reasoning models (e.g., GPT-5) achieve safety through constructive expansion (adding nuance), whereas capacity-constrained models revert to destructive truncation (deleting text). Finally, we identify a Reasoning Paradox: while inference-time reasoning increases baseline risk by enabling the model to make deeper sensitive inferences, it simultaneously empowers the defense to execute safe rewrites.

Setup

Environment Setup with UV

1. Install UV if not already installed:

   pip install uv
   

2. Initialize and activate the UV environment:

   uv venv
   uv pip install -r requirements.txt
   

   Or if using pyproject.toml:

   uv sync
   

3. Activate the environment:

   .venv\Scripts\activate  # On Windows
   source .venv/bin/activate  # On Unix
   

Repository Structure

• code/: Contains the main evaluation scripts, utility functions, and methods
• datasets/: Dataset files and processing scripts
• logs/: Log files from evaluations
• prompts/: Prompt templates used in the experiments
• visualizations/: Scripts and outputs for creating visualizations
• images/: Images used in the README and documentation
• semsi-datasets/: Additional SemSI-related datasets

Semantic sensitive information (SemSI)

Results

Acknowledgements

This work is heavily based on the original SemSI framework. We partially use their code and definition of Semantic Sensitive Information (SemSI).

• Repository: https://github.com/qingjiesjtu/SemSI
• Paper: https://openreview.net/forum?id=p3mxzKmuZy