Databricks Launches Agent-Based Security Review System in Under Two Hours

Databricks has developed a platform for agent-based security reviews, enabling the creation of a comprehensive system in under two hours. This system significantly reduces the time for eligible routine requests from days to minutes, enhancing efficiency in security assessments.

The architecture includes various specialized agents: the Intake agent identifies review paths and assembles structured requests, the Risk assessment agent assigns risk tiers based on evidence, and the Requirements agent maps requests to relevant standards. Additionally, the Validation agent builds checklists for higher-risk requests, while the Workflow agent manages follow-up tasks and escalations. The Learning agent improves the system by comparing reviewer edits to refine prompts and standards.

Databricks utilizes its Unity Catalog to govern security standards and manage request data, while Lakeflow Jobs orchestrates notebook-based workflows on serverless compute. The system leverages three Databricks-hosted foundation models: Claude Haiku for lightweight classification, Sonnet for most review tasks, and Opus for heavier reasoning requirements.

The automation rate is not specified, but the design aims to automate repeatable tasks while preserving human judgment for higher-risk scenarios, as emphasized by the author. This follows previous developments in Databricks' offerings, such as the recent general availability of Genie One MCP for AI agents. The system is intended to enhance expert attention where it is most valuable, rather than replace existing processes or personnel.