Particle.news
Download on the App Store

Databricks Expands Unity AI Gateway With Guardrails, Cost Controls and Tool Audit Logs

The company says the release will stop destructive agent actions, capture every external tool call in queryable logs, and let teams set granular AI budgets.

Overview

  • Databricks announced the expanded Unity AI Gateway on Tuesday, May 19, 2026, and made LLM-based guardrails plus AI Spend Controls available in beta while offering payload logging and MCP service policies through gated beta.
  • LLM guardrails run model-powered checks on inputs and outputs to block or sanitize sensitive prompts, and teams can tune evaluator models and prompts to balance precision, recall, cost, and latency.
  • Unity Catalog now governs Managed Cloud Platform tools by attaching SQL-defined service policies that evaluate each tool call by actor and context and return allow or deny decisions in real time.
  • Payload logging writes full request and response data into Delta-managed inference tables in Unity Catalog, including blocked requests and per-request token and DBU cost attribution for audits and debugging.
  • AI Spend Controls integrate with Databricks budgets to set per-user, per-workspace, per-use-case, and account thresholds with proactive alerts, and Databricks recommends gated beta enrollment and in-production tuning with account teams.