Databricks curbs AI agent dangers

Databricks Unity Catalog introduces granular control and logging for AI agents, preventing unauthorized actions and providing a complete audit trail.

May 19 at 7:02 PM8 min read

Databricks enhances AI agent security with Unity Catalog.

Visual TL;DR. AI agent dangers leads to Unrestricted tool access. Unrestricted tool access leads to Invisible agent actions. Invisible agent actions addresses with Databricks Unity Catalog. Databricks Unity Catalog enables Granular control. Databricks Unity Catalog provides Complete audit trail. Granular control leads to Curbs AI agent dangers. Complete audit trail enables Curbs AI agent dangers.

AI agent dangers: autonomous agents making critical decisions in production
Unrestricted tool access: agents had access to all available functions without fine-grained control
Invisible agent actions: tool calls were absent from standard logs, leaving no audit trail
Databricks Unity Catalog: extends governance to all connected tools, mirroring data management
Granular control: restricting write access or limiting administrative functions to specific users
Complete audit trail: providing a complete record of all agent actions and tool calls
Curbs AI agent dangers: preventing unauthorized actions and providing a complete audit trail

Visual TL;DRQuickExplainDeeper

The era of autonomous AI agents making critical decisions in production is here, bringing with it significant risks. Incidents of agents destructively wiping databases or deleting vast amounts of data are no longer theoretical, even when agents operate within their granted permissions. The core issue: a lack of control over which tools agents can access and no record of their actions.

Databricks is addressing this gap with new capabilities in its Unity Catalog, aimed at securing agent actions. The platform now extends governance to all connected tools, mirroring how data itself is managed.

The problem: Unrestricted access, invisible actions

Traditionally, if an AI agent was authorized to connect to a toolset, it had access to all available functions. This meant no fine-grained control, like restricting write access or limiting administrative functions to specific users. Crucially, when things went wrong, there was no audit trail.

Tool calls were absent from standard logs, leaving no record of what an agent did, with what parameters, or on whose behalf. This blindness makes it impossible to prevent unintended consequences or investigate incidents after they occur.

The solution: Unity Catalog MCP governance

Databricks is introducing granular control and comprehensive logging for AI agents through Unity Catalog MCP governance. This allows administrators to define precise rules for how agents interact with external tools, known as Managed Cloud Platform (MCP) tools.

Service policies, written in SQL, enable administrators to specify exactly which tools an agent can call and under what conditions. These policies can evaluate arguments and context, blocking or requiring consent for specific actions.

This is a significant step towards agentic AI security, preventing unauthorized agent behavior.

How it works

Administrators define service policies as SQL functions within Unity Catalog. These functions take the caller (actor) and the requested action (context) as input, returning an allow or deny decision with a reason.

For example, a policy could block all file deletion requests or permit merging pull requests only for approved engineers.

Once defined, these policies are attached to MCP services via the Databricks AI Gateway service policies. The gateway enforces these rules in real time for every tool call.

Simultaneously, all tool interactions are captured as entries in a Delta table within Unity Catalog. This includes the tool name, arguments, result, user identity, and whether the call was allowed or denied, providing a complete, queryable audit log.

This capability is currently available in gated beta, extending existing data governance principles to the realm of AI agents.

© 2026 StartupHub.ai. All rights reserved. Do not enter, scrape, copy, reproduce, or republish this article in whole or in part. Use as input to AI training, fine-tuning, retrieval-augmented generation, or any machine-learning system is prohibited without written license. Substantially-similar derivative works will be pursued to the fullest extent of applicable copyright, database, and computer-misuse laws. See our terms.

#Databricks #Unity Catalog #AI Security #Agentic AI #LLM Ops #Data Governance #Cloud Computing