Databricks enhances guardrails with sensitive data detection and conversation-aware LLM policies
Databricks has introduced a Sensitive Data Detection guardrail (Beta) that automatically identifies and blocks or redacts structured sensitive data like Social Security numbers or credit card numbers from requests and responses. This deterministic detection adds minimal latency and improves data privacy for applications. The update also enhances blocked interaction responses to prevent re-triggering, allows LLM-as-a-judge guardrails to evaluate entire conversation histories for better context, and provides a new execution flow view for guardrail operations. Users can enable the feature via the Enhanced Unity Gateway preview.
- →Sensitive Data Detection Guardrail (Beta)
- →New Guardrail Execution Flow View
- →Improved HTTP Responses for Blocked Interactions
- →LLM-as-a-Judge Guardrails Support Conversation History
Features (2) ›
- Sensitive Data Detection Guardrail (Beta)
A new guardrail automatically detects structured sensitive data, such as Social Security numbers and credit card numbers, to either block interactions or redact matched values. This deterministic detection uses regular expressions, checksums, and context keywords, adding under 50 ms for a 100-turn conversation. To use it, enable the Enhanced Unity Gateway preview.
- New Guardrail Execution Flow View
A new execution flow view has been added, which visually demonstrates the end-to-end order in which guardrails run during processing.
Enhancements (2) ›
- Improved HTTP Responses for Blocked Interactions
Blocked interactions now return a standard HTTP 200 response with the block reason, instead of a 4xx error. This improvement prevents coding agents from re-triggering the same block when resending conversation history.
- LLM-as-a-Judge Guardrails Support Conversation History
Built-in LLM-as-a-judge guardrails can now evaluate a window of recent conversation turns or the entire conversation history, providing more context than only the latest message for moderation.
https://docs.databricks.com/aws/en/release-notes/product/2026/august#block-or-redact-sensitive-data-in-requests-and-responses
Related releases
- Databricks Unity Gateway API now supports programmatic resource management Databricks Release Notes ·
- Schema-level HTTP connections in Databricks Unity Catalog are now GA Databricks Release Notes ·
- Databricks-managed MCP Connectors Integrate with Unity Gateway in Beta Databricks Release Notes ·
- Databricks-managed MCP Connectors Integrated with Unity Gateway (Beta) Databricks Release Notes ·
- Databricks Unity Gateway Achieves General Availability for Enterprise AI Governance Databricks Release Notes ·
- Unity Gateway API: Manage resources with SDKs, CLI, and Terraform Databricks Release Notes ·