NobleCloak
Category

Your OAuth grants are your first AI policy draft

Author

Audrey

NobleCloak's AI Correspondent · AI-drafted, fact-checked against our sourced evidence before publishing.

Date Published

Most AI policies start with a blank page and a list of worries. Start with a list of facts instead. Your identity provider already records every third-party app your people have connected to your Google Workspace or Microsoft 365, what each one can reach, and who said yes. Read that list and you have the first draft of your AI policy. Each connected tool is a decision someone already made about who may act and who must ask. Your job is to make those decisions on purpose.

Shadow AI lives in what people granted

For twenty years, shadow IT was a network question. Someone used a tool you hadn't approved, and you found out by watching traffic: firewall logs, a CASB alert, DLP catching a sensitive string on its way out.

Shadow AI mostly doesn't show up there. Picture a staff member who connects an AI assistant to their work account. A screen asks for permission to read their email, see and download their Drive files, and read their calendar. They click Allow.

Your firewall sees a connection to Google and to a mainstream AI vendor, both on the allowlist. No policy fires, because nothing sensitive moved in that moment. What changed is that an outside company now holds a standing grant to that person's mailbox and files. It can use that grant from its own servers, any time, without crossing your perimeter again.

That's why the inventory that answers "what AI do we use?" is a grant inventory. It's the list of every third-party app your people have authorized, the scopes each one holds, and who approved it. It maps straight to what an examiner asks about, because it describes access.

What an OAuth grant is

When an app says "Sign in with Google" or "Connect your Microsoft account," it uses OAuth to request scopes: named permissions such as "read your calendar" or "read and write your Drive files." If the person (or an admin) consents, the app gets a token it can keep using until someone revokes the grant. Logging out doesn't end it. A meeting assistant a former employee connected in March can still be reading calendars in July.

Two things make a grant risky, and both are on the console:

  • Scope: how much it can reach. Calendar read is one thing. Read and write access to all of Drive or all of Mail is another.
  • Breadth: how many people it covers. A grant one person made for themselves is contained. A grant an admin made for the whole organization exposes everyone through one app.

Pull the list: Google Workspace

You need the Service Settings admin privilege for the controls and Audit & Investigation for the logs.

  • Who has what now. In the Admin console, go to Security > Access and data control > API controls, then Manage App Access (Google). Under Accessed apps, click View list to see every app that has reached Google data, how many users it covers, and the services or OAuth scopes it used. Each app can be set to Trusted, Limited, Specific Google data, or Blocked.
  • When it happened. Go to Reporting > Audit and investigation > OAuth log events (Google). It records each time a third-party app was authorized to reach account data such as Contacts, Calendar and Drive, with the user.

Pull the list: Microsoft Entra ID

You need at least the Cloud Application Administrator or Application Administrator role.

  • Who has what now. In the Microsoft Entra admin center, go to Entra ID > Enterprise apps > All applications, open an app, and select Permissions (Microsoft). The Admin consent tab shows permissions granted for the whole organization. The User consent tab shows what individual people granted.
  • Read the admin-consent tab first. Microsoft's own documentation says "Granting admin consent on behalf of an organization is a sensitive operation, potentially allowing the application's publisher access to significant portions of the organization's data" (Microsoft).
  • When it happened. Entra's audit logs record consent events, and sign-in logs show the app's activity.

Read each grant

Here's a typical find, written the way you'd note it:

An AI meeting-notes app holds Calendar and Drive read access for 8 users since March. Recordings are kept on the vendor's side. No security attestation on file.

  • The app: an outside company holding tokens to your tenant.
  • Calendar and Drive read: who meets whom and about what, plus every file those eight people can open.
  • 8 users since March: eight standing connections, live for months.
  • Recordings kept, no attestation: the two facts the console can't tell you. You add them from vendor diligence.

Then ask three questions, in order:

  • What can it reach? Mail and full Drive access are high. Calendar read and basic profile are lower. Write access outranks read-only.
  • How many people does it cover? An organization-wide grant, or one held by many users, outranks a single-user grant.
  • What does the vendor do with the data? A tool that keeps or trains on what it reads outranks a well-known vendor with a clean data agreement.

Turn the list into your first policy draft

Every row on that list is a tool, and every tool needs one of three decisions:

  • Allowed. The tool is vetted, its scopes fit the job, and the data it reaches is low-risk. Record why, and when you decided.
  • Needs approval. The tool is useful, but some of what it can do should wait for a person. Narrow its scopes where the console allows (Google's Specific Google data setting, or user assignment in Entra), and decide who approves the risky actions.
  • Denied. High reach, no diligence, or no business need. Block it in Google or revoke the consent in Entra. In Entra, user-consent grants can't be revoked in the portal; that takes Microsoft Graph or PowerShell.

Write those decisions down per tool and you have a policy grounded in what your people actually use. It's also the shape an examiner can follow: here is what's connected, here is what we decided for each, and here is the date.

Two settings make the draft stick. In Entra, the admin consent workflow sends people an "Approval required" message when they try to connect an app they can't consent to on their own. In Google, apps you haven't trusted can be limited or blocked. Either way, the next new tool becomes a decision instead of a surprise.

Why this is a compliance question too

If you're an RIA or broker-dealer, the amended Reg S-P rule (17 CFR 248.30(a)(5)) requires ongoing oversight and monitoring of service providers. An AI app holding a standing grant to your systems has access to your data, contract or not. The grant list is how you'd know it exists to oversee. For RIAs, the next step is the Reg S-P vendor file.

If you're a credit union, NCUA's letter on third-party relationships (07-CU-13) says that, at a minimum, due diligence "should take into account the critical nature of the service." An app with a standing grant to member data is a service to weigh that way, and the grant list is how you find it. The diligence questions themselves are in AI vendor risk for a 1–2 person compliance program.

What a grant inventory doesn't show

It shows access. A grant tells you an app can read your Drive. It doesn't tell you what the app read, or what it did with it. For some tools you can't get that record at all on the plan you're paying for; the ChatGPT compliance log export, for one, is limited to Enterprise and Edu (details).

It's a point-in-time read. People connect new tools every week, so the value is in re-running it and comparing.

And it misses whatever doesn't go through your identity provider: someone pasting member data into a personal chatbot on their phone, an AI feature a vendor runs inside its own product, or a tool that connects with an API key instead of OAuth. Network and DLP controls still earn their keep for those.

Where NobleCloak fits

Our read-only scan reads Google Workspace and Entra and returns every AI app your people authorized, the data scopes it holds, who's using it, and since when. Coverage is what SSO and OAuth can see, and the report says so. Our team runs the first report with you. See the AI Trust Report.

Coming soon: connecting the scan yourself, mapping the apps it finds to tools NobleCloak can govern, and drafting your policy set from the scan. Today, the three decisions above are yours to make, with the list in hand.

When an agent does act through one of those tools, NobleCloak's policy marks each tool as allowed, needs approval, or denied, and the Record of Decisions keeps who asked, which tool, and who approved. Denying OAuth consent to AI outright is one of the IT and security examples.