Policy Once, Enforced Everywhere: Alice WonderFence Joins Databricks Unity AI Gateway

The Data + AI Summit has been packed this week with sessions and talks about building agents. But what about controlling them in production, and making sure what you built actually serves your brand?

Today, Databricks announced the Unity AI Gateway partner ecosystem, opening its enforcement layer for the agentic stack to the security and governance tools enterprises already trust. Alice is one of the launch partners, named for runtime AI security and guardrails. WonderFence, our real-time guardrails platform that can be trained on your own policies, plugs directly into the gateway's enforcement path: detectors tailor-made for your brand, your users, and your regulatory environment. Your policies stop living next to your AI workflows and start living inside them.

Here is what that means in practice, and why it changes how security and safety teams work on Databricks.

‍

One policy, attached to everything

Unity AI Gateway introduces Service Policies: a centralized way to define how AI workloads are allowed to behave. Service Policy can govern model requests, MCP tool calls, and agent interactions, and travels with the workload across Databricks workspaces.

The integration connects those policies to the WonderFence platform, where organizations can define custom AI safety and security controls tailored to their specific use cases. These include protections against prompt injection and jailbreaks, PII handling, off-brand or competitor responses, and other unwanted content across text, image, audio, and video modalities. Service Policies reference these controls, and the gateway enforces them in real time.

The flow is simple:

→ A request enters the Unity AI Gateway

→ The gateway checks the attached Service Policy

→ WonderFence evaluates the request against the organization’s policies in real time

→ The request is allowed or blocked before it reaches the model, and the same enforcement is applied again before the response reaches the user

‍

From the agent's code to the platform's spine

This announcement deepens our ongoing partnership with Databricks, devoted to helping enterprises build fleets of agents that are safe, secure, and on track with business goals.

Alice and Databricks first brought WonderFence into the Mosaic AI Agent Framework, adding runtime protection to agents through the Alice SDK, citing real-time protection of agents across 117+ languages. That integration asked developers to build safety into each agent. Our new one builds safety into the platform every agent runs on.

The difference shows as agent fleets grow. Ten agents wrapped in guardrail code is an engineering pattern. Two hundred agents governed by policies enforced at the gateway, managed in one place and consistent across every workspace, is an operating model that can scale: the next hundred agents inherit the same policies, the same enforcement, and the same audit trail on day one.

"Alice is thrilled to help organizations building on Databricks govern AI workflows with personalized runtime guardrails. Integrating Alice's WonderFence with Unity AI Gateway delivers consistent policy enforcement across Databricks workspaces, so our shared customers can advance unafraid as they deploy AI at scale."

_{– Avi Golan, Chief Product & Engineering Officer, Alice}

‍

Where out-of-the-box guardrails reach their limit

The guardrails that ship with foundation models handle broad, well-understood harms well. The cases that tend to slip through are the ones shaped by your specific context, and by the people actively working to break your system.

On the security side: multi-turn prompt injections that walk a support bot into generating profanity against its own brand, jailbreaks that exfiltrate your organization's PII, or reverse engineering tricks that drink your tokens.

‍

Read more about 5 Ways to Break Your Chatbot.

‍

And beyond security, no generic model carries regulatory nuance: where information ends and a recommendation begins in healthcare, or what an assistant may and may not say about financial products, where that nuance is the entire compliance question.

This is the gap policy-trained guardrails close. Teams define the policy, provide examples of acceptable and unacceptable behavior, and WonderFence trains detectors against real adversarial data.

–

Enterprise AI is stepping into a new era of governance and observability, and we are glad to be building it alongside Databricks. With Unity AI Gateway, those detectors sit in the same enforcement path as the rest of your governance: define the policy once, attach it to your models, tools, and agents, and get back to building with WonderFence.

‍