verifiedagents.ai
All posts

5 min readAI Agents · Cybersecurity

How do you test your AI agent inventory against your OAuth consent grants?

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

Hero: How do you test your AI agent inventory against your OAuth consent grants?

TL;DR: Export every app your people clicked "allow" for, then match that export against your AI agent inventory. Each AI app with a grant and no inventory row is a miss. The test takes one person about 30 minutes and needs no new product. Pass means zero misses and an owner on every row.

Every time someone clicked "allow" to connect an AI app to their email or files, your identity provider wrote it down. That record is called an OAuth consent grant. It's the one inventory nobody had to remember to fill in.

So it makes a good answer key. Your agent list is the test paper.

It proves who approved an app and what that app can reach. It's created the moment of the click, whether IT approved the app or not.

That click is the whole control event. Someone connects a note-taker to a calendar, and the app now runs with that person's access until somebody revokes it.

The longer explainer is on MassiveScale.AI, in what an OAuth consent grant is and why it's your fastest AI inventory. Here I'm using the grants to grade the list you already have.

How does the assessment score this?

Question 1 of the free assessment asks how you identify AI agents in your organization. It belongs to Identity Management, the first of the five ATF elements. The grant export tests each of its five answers in a different way.

Answer

What it says

What the grant match shows

A

We have no formal process. We're not sure which agents exist.

No list to match. The export becomes your first one.

B

We track agents informally, in spreadsheets or team knowledge.

Many AI apps with grants and no row anywhere.

C

We keep a basic inventory that lists each agent with its purpose and owner.

Most apps match. The misses come from personal sign-ups.

D

Each agent has a unique ID in one trusted system, with its permissions written down.

Each grant maps to an ID, and the scopes match what was written.

E

Every agent is registered with cryptographic credentials and an ownership chain, plus a record of what it can do.

People can't approve apps on their own, so new grants are rare.

How do you run the match in Microsoft Entra?

I've run this with three clients between June and September 2026. One was a global manufacturer and one was a big-box retailer. The other was a two-person security team at a therapeutics company, who did it live on one call. The screen with the answer had been open the whole time.

  1. Open Enterprise applications. In the Entra admin center, go to All applications. Every app anyone approved is a row.

  2. Filter and sort. Set Application type to Enterprise Applications and sort by created date. The newest rows come first.

  3. Check permissions on the AI apps. Open each one and click Permissions. The user consent tab shows who approved it and what it reaches.

  4. Export and match. Put the export next to your inventory. Mark every AI app that has no inventory row.

  5. Count. The marked rows are your misses.

Microsoft documents the same path, plus PowerShell and Graph API versions, in its guide to reviewing permissions granted to enterprise applications.

What's the pass bar?

Two things have to be true.

  • Zero misses. Every AI app with a grant has a row in your inventory.

  • Every matched row names an owner.

Then sort the misses by reach. An app that can write and send goes to the top of the pile. Read-only apps come after.

Where do the records live outside Entra?

Google Workspace and Okta keep the same kind of record.

Provider

Where to look

What to watch for

Google Workspace

Admin console, Security, API controls

Domain-wide delegation, where an app can act as any user

Okta

System Log, token grant events

Which app acted for which person

What should you do with the misses?

Close the front door first. In Entra, change user consent so people can only approve apps from verified publishers, or so an admin approves each one. Turn on the admin consent workflow so new requests go to a named reviewer.

Then add the misses to the inventory or revoke them. Tell each person before you revoke. People who needed the app will ask for it back. People who forgot it existed won't.

One limit to say plainly. This test only sees apps connected through your identity provider. An app used in a browser with a personal email leaves no trace here. That's why it's one of four sources in checking whether your AI agent inventory is complete.

Frequently asked questions

Do I need to buy a product for this?

No. Your identity provider already keeps these records, and the admin screens come with what you pay for today.

Will revoking a grant break something?

It can. Tell the person first and publish the approved path before you prune.

What's the difference between a delegated permission and an app permission?

A delegated permission is what one person approved for themselves. An app permission runs on its own, tied to nobody's click. Both show up in Entra under different tabs.

What do I do with apps that pass?

Check that their access fits the job. That's the next test, in whether an AI agent has more access than its job needs.

Key takeaways

  • Every "allow" click left a record you can read today.

  • The grant export is the answer key, and your inventory is the test paper.

  • A miss is an AI app with a grant and no inventory row.

  • Pass means zero misses and an owner on every matched row.

  • The test can't see apps used outside your identity provider.

Question 1 is one of 30 in the free assessment. It takes about ten minutes and scores you on all five ATF elements.

The screen with the answer has been open the whole time. Put your list next to it and count.

See where your agents stand.

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