Summary

In its August 11, 2026 platform release, Databricks expanded Unity Catalog ABAC GRANT policies beyond models to cover model services, model provider services, Model Context Protocol (MCP) services, agent services, and skills (all in beta). Tag-conditioned policies now dynamically grant privileges across these AI assets, extending governance from data to the runtime interactions between models, agents, MCP services, and enterprise tools via the Unity AI Gateway.

What changed

Databricks broadened Unity Catalog ABAC GRANT policies to apply to model services, model provider services, MCP services, agent services, and skills, so tag-based conditions govern access to these AI assets (beta).

Why it matters

As enterprises deploy agents, MCP servers, and reusable skills, access control has to reach those runtime assets, not just tables and models; putting them under the same attribute-based policy engine gives platform teams one governance model for data and agent infrastructure, a prerequisite for compliant production agents.

Evidence excerpt

ABAC GRANT policies were previously only available for models, but now also support model services, model provider services, Model Context Protocol (MCP) services, agent services, and skills. All supported types are in Beta.

Sources