Policy-Aware SQL Blocks 80-Row Leaks With 6.5-Millisecond Overhead
Updated
Updated · InfoWorld · Nov 5
Policy-Aware SQL Blocks 80-Row Leaks With 6.5-Millisecond Overhead
1 articles · Updated · InfoWorld · Nov 5
Summary
900 benchmark queries showed a policy-aware LLM-to-SQL layer eliminated unauthorized data leakage while adding about 6.5 milliseconds per query, outperforming plain text-to-SQL and static role-based controls.
285 query rewrites added WHERE-clause row filters, 240 applied column masking, and 165 restricted scope, reflecting a design that checks user role, territory and assigned accounts before execution.
40% of queries ran unchanged, while 20% needed row filters, 15% needed masking, 11.7% required join rewrites, and 13.4% were either sent to human review or blocked.
The approach also rejects SELECT *, cross-scope pulls, oversized result sets and risky joins, then routes approved SQL through a narrow database account to contain prompt injection or bad generations.
The work, accepted at IEEE ICICNIS 2026 and tested on synthetic de-identified data, argues that access policy and audit trails matter more than raw SQL-generation accuracy for trustworthy analytics.