Using an LLM to translate questions into database queries—high leverage for exploration, dangerous without validation, access control, and query review.
NL→SQL pairs language understanding with your schema. Failures include wrong joins, subtle filter bugs, and over-broad scans. Security overlaps with classic injection risks when models emit dynamic SQL.
Controls
Read-only roles, row-level security, statement limits, human approval for writes, and golden test questions on representative schemas.