Skip to content

DASF - Databricks AI Security Framework

Preetham ReddyUpdated 2 min read

Key challenges in securing AI: an arrow from traditional risks to new AI-specific risks, beside four challenges: massive and diverse data, complex AI supply chain, dynamic model behaviour and threats, and regulatory compliance and ethical use.

DASF is the Databricks AI Security Framework. Version 3.0, published in March 2026, catalogs 97 technical security risks across 13 components of an AI system, from the raw data and the models through serving, agents and the platform underneath, and maps them to 73 controls. It's a useful map, but you don't install it. The work is deciding which of those risks you actually have, then putting a gate in the runtime.

That gate is the difference between a policy and a control. A document describing how an agent should behave won't stop it sending an email it shouldn't. A control stops the action when the grant is missing, and leaves a record either way.

Where we put the gates

Data. Unity Catalog decides who can see what. Agents and apps run under their own service principal, on the same path a person uses. If the catalogue coverage stops halfway, that's a Health Check finding, and it will show up in an agent programme later as an incident.

Models and tools. Unity Gateway is where model serving is governed. Tools an agent can call resolve through the catalogue, so it can't reach a table its principal has no grant for. TechFabric Harness is how we deploy that agent as a Databricks App without rewriting it for each target.

Actions. Every domain change goes through one mutation pipeline. TechFabric Platform makes an illegal transition structurally impossible. An agent actor passes the same policy and state-machine gates as a human one, and the event that explains the change is part of the same transaction.

Approvals. Anything that would leave the workspace parks at a ticket a person answers. Fabric Tower is the console. The supervisor can't approve itself.

no grant

stays in workspace

leaves workspace

Agent runs as service principal

Unity Catalog checks grant

Stopped, recorded

Domain change enters mutation pipeline

Policy and state-machine gates

Change and event commit together

Parks at ticket

A person answers

An agent passes the same catalogue, policy and state-machine gates as a person, and anything leaving the workspace waits at a ticket for a person to answer.

The DASF compendium already walks all thirteen components, and a paraphrase here would be worse than the source. The useful question is whether your agents pass the same gates as your people. A second, weaker path is the incident.

DASF is Databricks' framework, not a certification we hold. We use the surfaces it assumes: Unity Catalog, Unity Gateway, Model Serving, Apps. We don't claim a partner specialisation we don't have.

Almost every team that asks us about AI security has not yet written down what a good answer is. Build the rubric first. Genie Accuracy and TechFabric Experiments are that work. Then the gates have something real to protect.

The service line is AI systems. Troy's line still holds: draw the permissions boundary in the first conversation.

Frequently asked questions

What happens to an agent when its service principal has no grant for a table?

The tool call can't reach the table, because tools resolve through the catalogue under the agent's own principal. The action stops at the gate and leaves a record, which a policy document describing the agent's behaviour can't do.

Can an agent approve its own action if it's running as a supervisor?

No. Anything that would leave the workspace parks at a ticket that a person answers, and the supervisor agent has no way to clear that ticket itself. That also means an approval backlog is a people problem to staff, so someone needs to own the queue before the agent goes live.

Is DASF a certification we can point a customer to?

No. It's Databricks' framework, version 3.0 from March 2026, mapping 97 risks across 13 components to 73 controls, and it isn't a certification. What you can show is the gates: an agent passing the same catalogue, policy and approval checks as a person.