arXiv: Policy-Conditioned Constrained Decoding for Column-Level Access Control in Text-to-SQL

This publication introduces a technical framework called Policy-Conditioned Constrained Decoding, designed to enforce column-level access control in text-to-SQL systems. The paper proposes a method that ensures large language models generating SQL queries from natural language only access database columns permitted by predefined organizational policies. This addresses a critical gap where AI-generated queries could inadvertently expose sensitive data, such as personally identifiable information or financial records, even if the underlying database has row-level security.

The primary affected organizations are those deploying AI-powered natural language interfaces to databases, particularly in regulated sectors like finance, healthcare, and insurance. Compliance teams in these industries must now consider that existing data access controls may not automatically extend to AI-generated queries. The framework is especially relevant for firms subject to GDPR, HIPAA, or SOX, where unauthorized column access could constitute a data breach or compliance violation.

Compliance teams should immediately review their current text-to-SQL implementations to verify whether column-level access controls are enforced. They should assess whether their AI vendors or internal models incorporate similar constrained decoding techniques. Next, teams should update data governance policies to explicitly cover AI-generated queries and conduct a gap analysis between existing database permissions and the queries produced by these systems. Finally, consider piloting this framework or equivalent solutions to ensure audit trails and access logs capture all column-level interactions from AI tools.