SAFENLIDB: A Privacy-Preserving Safety Alignment Framework for LLM-based Natural Language Database Interfaces

The publication of SafeNlidb marks a significant advancement in addressing the privacy and security challenges associated with the growing use of Large Language Models (LLMs) in Natural Language Database Interfaces (NLIDB). As LLMs become more prevalent, they pose risks of unintentionally exposing sensitive database information or being exploited by malicious actors to extract data through innocuous queries. Current mitigation strategies often rely on rule-based heuristics or LLM agents, which can struggle against complex attacks and lead to high false positive rates. SafeNlidb proposes a novel approach that combines implicit security reasoning with SQL generation through an automated pipeline, effectively generating hybrid chain-of-thought interaction data. This innovative framework not only enhances security but also improves the reliability of SQL queries. Extensive experiments have demonstrated that SafeNlidb outperforms both larger-scale LLMs and ideal-setting baselines, achieving…