verifiedagents.ai
All posts

5 min readAI Agents · Zero Trust

How do you score your organization against the Agentic Trust Framework?

By Josh Woodruff, Founder & CEO, MassiveScale.AI | Founding Chair, Agentic Trust Framework at the CSAI Foundation

Hero: How do you score your organization against the Agentic Trust Framework?

TL;DR: Score your organization against the Agentic Trust Framework by answering 30 questions, six for each of its five elements. Each question has five answers, from no control to a fully automated one. Only claim an answer you can show evidence for. Your lowest element score tells you where to start.

I created the Agentic Trust Framework (ATF) and donated it to the CSAI Foundation, the Cloud Security Alliance's nonprofit for AI agent accountability. It was published in February 2026 under a CC BY 4.0 license. Anyone can read the spec at agentictrustframework.ai.

The question I get most isn't what the framework says. It's this one. How do we know where we stand against it?

You score yourself. Here's how to do it so the number means something.

What does the framework measure?

Five elements. Each one asks a plain question about your agents.

Element

The question it asks

Identity Management

Who are your agents?

Behavioral Monitoring

What are they doing?

Data Governance

What data flows through them?

Segmentation

Where can they go?

Incident Response

What if they go rogue?

Zero Trust is the foundation under all five. Never trust, always verify. The ATF adds one more check on top, at the level of each action an agent takes. What the framework is and why it exists is on MassiveScale.AI, in what the Agentic Trust Framework is.

How is the score built?

The free assessment asks 30 questions. That's six for each element.

Every question has five answers. They run from the weakest to the strongest.

Answer level

What it usually means

Example, from the kill switch question

First

Nothing is in place.

No ability to stop agents instantly.

Second

A general practice covers it loosely.

We can restart or shut down through normal operations.

Third

It's written down.

We have a documented emergency shutdown process.

Fourth

It's enforced by a control.

We have a one-click kill switch.

Fifth

It's automated and adapts.

Automated kill switches triggered by anomalies.

You get a score for each of the five elements and a picture of where you're weak. Then you get a PDF that says what to fix first.

How do you keep the score honest?

Four rules. I use all of them when I score a client.

  1. Answer for your typical agent. Don't answer for your best one. If one agent has a kill switch and forty don't, you don't have kill switches.

  2. Claim only what you can show. If you pick the fourth answer, you should be able to open the control. A plan to build it earns the answer below.

  3. Ask the person who runs it. The team that operates the agents knows the real answer. Leadership often knows the intended one.

  4. Write down one piece of evidence per answer. A link or a file name is enough. You'll want it later.

Rule 2 does the most work. Most high scores I've had to correct came from counting a policy as a control.

What counts as evidence for an answer?

It depends on the level you're claiming.

If you claim

You should be able to show

The second level (a general practice)

The existing process you're relying on, by name

The third level (written down)

The document, with a date and an owner

The fourth level (enforced)

The control itself, and a log entry of it working

The fifth level (automated)

A record of it firing with no person involved

I've mapped the evidence to an auditor's questions in what evidence passes each of the five AI agent audit questions.

What do you do with the result?

Read the lowest element first. Skip the total for now.

A chain breaks at its weakest link. Strong Identity Management doesn't help much if Incident Response is at the first level and nobody can stop an agent.

Then pick one question in that element and move it up one level. One level is the right-sized goal. Going from "written down" to "enforced" on a single question is real progress you can show.

Two posts help with the next step. To decide how much freedom a new agent gets, use how to score an AI agent approval request. To move an existing agent up, use what evidence proves an AI agent is ready to move up a level.

How should you talk about your score?

Carefully. The words carry claims.

"Aligned with ATF" and "maps to ATF" are safe ways to describe work you've done. "ATF-conformant" means a real conformance assessment happened against all 25 requirements. A self-score isn't that.

The ATF is voluntary, so there's nothing to comply with. The word is conformant.

Other teams are doing this mapping in public. Microsoft's Agent Governance Toolkit lists all five ATF elements as mapped in its own documentation.

Frequently asked questions

How long does scoring take?

About ten minutes for a first pass. Plan another hour to collect one piece of evidence per answer.

Who should take the assessment?

Someone who knows how the agents really run. For a bigger organization, have each team answer for the parts it owns and compare.

How often should we rescore?

Every quarter. Rescore after any big change to your agent platform too.

Is the ATF only for big companies?

No. The controls you need depend on what your agents do. Headcount doesn't decide it.

Key takeaways

  • The ATF has five elements, and the assessment asks six questions for each.

  • Each question has five answer levels, from nothing in place to automated.

  • Claim only the answers you can show evidence for.

  • Start with your lowest element and move one question up one level.

  • A self-score isn't conformance. Say "aligned with ATF."

Take the free assessment. It's 30 questions and about ten minutes.

Your agents are making decisions right now. A score won't change that, but it'll tell you which decision to worry about first.

See where your agents stand.

The free assessment takes ten minutes and scores you on the five elements of the Agentic Trust Framework.