One Developer's Gatekeeper Pattern for Locking Down AI Agent Database Access

A software developer shared a governance pattern after an LLM agent nearly executed a DROP command on a live Azure Data Explorer cluster. The approach introduces a gatekeeper layer that intercepts every SQL query from an AI agent before it reaches the database, blocking anything that isn't a SELECT statement. A strict blocklist of destructive verbs — including DROP, DELETE, and ALTER — is enforced by default, with unknown commands also rejected. Every query verdict, whether allowed or blocked, is recorded in an append-only immutable log the agent cannot modify. The author implemented this pattern in three ways: using Open Policy Agent with Terraform for production, client-side JavaScript for browser demos, and notes that the policy layer complements rather than replaces read-only database replicas.