Foundation Model governance expanded with Unity Catalog permission controls and ABAC grant policies.

In September 2026, governance controls around hosted model serving became a primary focus. Databricks enabled granular Unity Catalog permissions specifically for Foundation Models [3]. Organizations also received technical guidance on governing models and model serving endpoints through attribute-based access control (ABAC) grant policies [5]. This configuration used catalog tags paired with programmatic exclusion statements to restrict access and replace broad schema-level grants [5].

Foundation Model APIs increasingly served as the backend for structured agent deployments. One pattern detailed an automated security review architecture that paired hosted foundation models with Unity Catalog, Lakeflow Jobs, and Databricks Apps to handle routine evaluations while escalating edge cases to human reviewers [1]. External agent frameworks like Omnigent also integrated Databricks Foundation Model APIs to route model calls across different execution harnesses, keeping cost tracking and access rules centralized under one control plane [2].

Community discussions centered on operationalizing these endpoints. Developers sought methods to inject runtime safety guardrails into existing LangGraph agents running on Databricks Model Serving [6]. Meanwhile, community sessions highlighted production architectures combining streaming data pipelines with foundation models for real-time applications [4].