Text-to-SQL Benchmarking Under Role-Based Access Control

A new benchmarking framework addresses the gap between text-to-SQL performance on standard benchmarks and real-world deployment under role-based access control (RBAC). Current benchmarks assume unrestricted database access, but in practice user queries are limited by RBAC policies. The proposed framework augments existing benchmarks with plausible user roles and access policies using an LLM-assisted workflow. It formulates role synthesis as a structured problem to generate realistic constraints. This work aims to evaluate how well LLMs handle SQL generation when data access is restricted, highlighting potential failures in both policy violations and missed permissible queries.