Locking Down Copilot's Reach: Restricted Content Discovery, Agent Governance, and Work-Only Access

Share
Locking Down Copilot's Reach: Restricted Content Discovery, Agent Governance, and Work-Only Access
Modern Endpoint · Security Insights

Locking Down Copilot's Reach: Restricted Content Discovery, Agent Governance, and Work-Only Access

I walked into a Copilot readiness review last quarter where the security team was genuinely proud. They had completed their sensitivity label taxonomy, trained their users on data classification, and built a solid Conditional Access baseline. The CISO asked me directly: "Are we ready to roll out Copilot?" I told her the labels were excellent. Then I asked who had access to the legal department's SharePoint site. She looked at her SharePoint admin. He looked at his screen. Thirty seconds later, we found that 4,200 employees had read access to a site containing merger-related documents, grandfathered from a company-wide "everyone reads" policy set in 2019. Nobody had touched it since. The labels were fine. The permissions were the problem. And Copilot was about to read all of it at once.

11 min read ArticleModernEndpoint

🔍 The Real Exposure Surface

SharePoint permissions have always been imperfect. Sites get created with overly broad access. Folders get shared with "everyone in the organization" because someone needed a quick fix in 2021. Inherited permissions accumulate across years of reorganizations. Security teams knew this. Most of them had accepted it as low-risk operational debt, because in practice, a human browsing SharePoint can only read what they actively navigate to. The surface area, while technically wide, was practically narrow. A user has to know a file exists to find it.

Microsoft 365 Copilot collapses that distinction completely. Copilot searches and synthesizes across everything a user has access to, across SharePoint, OneDrive, Teams, Exchange, and more, in seconds. It doesn't browse. It retrieves. A prompt like "summarize any documents about Project Falcon" doesn't require the user to know which site the document lives on. Copilot finds it, pulls it, and summarizes it. The technical access was always there. The practical barrier was human friction. Copilot removes the friction.

That is the governance gap nobody warned the CISO about.

⚡ Assumption Challenge
Most organizations believe: "Our SharePoint permissions are already in place — Copilot just respects them."
Reality: Copilot does respect permissions, exactly as defined. The problem is that permissions defined over eight years of organic growth were never designed to survive AI-scale retrieval. Respecting bad permissions at machine speed is worse than ignoring them at human speed.

⚠️ Three Levers That Actually Matter

There are three concrete controls that address this exposure. Not awareness campaigns. Not generic "fix your permissions" advice. Actual architectural levers.

Lever 1: SharePoint Advanced Management and Restricted Content Discovery

SharePoint Advanced Management (SAM) is the feature set most organizations overlook when they plan Copilot governance. It includes Restricted Content Discovery, which lets administrators mark specific SharePoint sites so that their content does not surface in Copilot responses or Microsoft Search results, even when a user technically has access to the content.

This is not a permissions change. The site stays accessible. A user who navigates directly to the URL can still read the content. What changes is that Copilot's semantic search index treats the site as out of scope. Think of it as a fence around the discovery surface rather than a lock on the door.

🔍 Reality Check
What most organizations believe: Sensitivity labels on documents are enough to control what Copilot surfaces.
What actually happens in production: Copilot respects sensitivity labels for DLP enforcement, but it still surfaces content from any site the user can access unless Restricted Content Discovery is explicitly configured. A "Confidential" label on a document does not prevent Copilot from summarizing it for a user who has read access to the site. Labels and discovery scoping are separate controls — both must be configured.

I strongly recommend organizations run a data access governance report inside SharePoint Advanced Management before they touch any Copilot deployment settings. The report shows overshared sites, sites with broad access, and "everyone" links. In most tenants I've assessed, the list is longer than the security team expects.

You can enable Restricted Content Discovery on a site using PowerShell:

powershell
Connect-SPOService -Url https://yourtenant-admin.sharepoint.com

Set-SPOSite -Identity "https://yourtenant.sharepoint.com/sites/LegalMergers" -RestrictContentOrgWideSearch $true

Documentation for this capability lives at: https://learn.microsoft.com/en-us/sharepoint/restricted-content-discovery

Warning

Restricted Content Discovery does not retroactively remove content already indexed in Microsoft Search. There is a propagation delay, typically measured in hours. Do not treat it as an immediate kill switch for sensitive content. Plan ahead.

Lever 2: Copilot Studio Agent Creation Governance

Most organizations are so focused on Microsoft 365 Copilot that they miss the adjacent risk entirely. Copilot Studio allows users, not just developers, not just IT admins, any licensed user, to build agents that connect to SharePoint libraries, external APIs, and organizational data sources. By default, users with a Copilot Studio license can create agents and share them across the organization.

I've seen this pattern emerge within 30 days of a Copilot license rollout. A well-meaning business analyst builds an agent connected to the HR SharePoint library, shares it with their team via Teams, and suddenly a dozen people are querying employee compensation data through an unsanctioned agent with no DLP inspection, no audit trail that the security team monitors, and no governance approval.

⚡ Assumption Challenge
Most organizations believe: "Copilot Studio is a developer tool — our users won't build agents without IT involvement."
Reality: Copilot Studio is designed for citizen developers. Its no-code interface means a user with business context and a SharePoint connection can build a functional data-querying agent in under an hour. The barrier is not technical. It's governance — and most organizations have none when they first go live.

The governance control here lives in the Power Platform admin center. Administrators can restrict who can create Copilot Studio agents by configuring environment-level security and by using the Microsoft 365 admin center to control which users have access to Copilot Studio authoring capabilities. Agent sharing can also be restricted so agents created by users cannot be distributed outside a defined group.

If I were designing this for an enterprise today, I would set the default posture to: no user-created agents published to the organization without explicit IT review and approval. Agents for personal productivity within a user's own context carry lower risk. Agents shared broadly against organizational data sources carry substantial risk and must go through a formal review cycle.

Note

Controlling Copilot Studio agent creation requires attention to both the Microsoft 365 admin center settings and the Power Platform admin center environment policies. These are separate administrative surfaces. A change in one does not automatically propagate to the other. Most organizations configure one and assume they've covered both.

⚖️ Trade-Off
Restricting Copilot Studio agent creation stops unauthorized data exposure. It also blocks legitimate innovation. Business units that want to automate workflows and build productivity tools will push back. The governance model must include a review and approval path, not just a block. Organizations that only block without offering an alternative channel create shadow AI use — users find external tools to do what Copilot Studio would have done, and now those tools are entirely outside the Microsoft security boundary. A restrictive posture without an approval pathway is not governance. It is friction that moves the risk somewhere harder to see.

Lever 3: Tenant Restrictions and Conditional Access for Work Account Isolation

The third exposure is architectural and often the least visible. Copilot is licensed to the Work account. But nothing inherently prevents a user on a personal device from signing into their Work Microsoft 365 account and running Copilot prompts against organizational data. More critically, nothing prevents a user on a managed corporate device from running Copilot through a personal Microsoft account if they navigate to copilot.microsoft.com and authenticate with a personal identity.

Tenant restrictions v2 combined with Conditional Access address both sides of this problem. Tenant restrictions v2 allows organizations to define which external Microsoft Entra tenants their users can authenticate against on managed network paths or managed devices. When combined with Conditional Access policies that require compliant, managed devices as a condition for accessing Microsoft 365 Copilot, the result is: Copilot runs only on managed devices, against the Work account, within the organizational tenant boundary.

🔍 Reality Check
What most organizations believe: Conditional Access controls who can access Microsoft 365. Copilot inherits those controls automatically.
What actually happens in production: Copilot at copilot.microsoft.com is a separate application endpoint from the Microsoft 365 portal. Organizations that have not explicitly scoped their Conditional Access policies to include the Copilot application IDs frequently discover gaps where Copilot is accessible from unmanaged devices that cannot reach other M365 workloads. The application scope matters. Assuming inheritance is how organizations create gaps they cannot see.

Documentation for tenant restrictions v2 is at: https://learn.microsoft.com/en-us/entra/identity/enterprise-apps/tenant-restrictions-v2

The Conditional Access policy to enforce managed device compliance for Copilot should target the Office 365 cloud app or specifically the Copilot application ID within the policy scope. Require device compliance as a grant control, not just as a session control. Session controls limit what happens during access. Compliance as a grant condition prevents access entirely from non-compliant devices.

json
{
  "displayName": "Require Managed Device for Copilot",
  "state": "enabled",
  "conditions": {
    "applications": {
      "includeApplications": ["Microsoft Copilot application ID"]
    },
    "users": {
      "includeGroups": ["Copilot-Licensed-Users"]
    }
  },
  "grantControls": {
    "operator": "AND",
    "builtInControls": ["compliantDevice"]
  }
}
🎯 Enterprise Decision Point
Before enabling Copilot for any user population, decide: which SharePoint sites should be excluded from Copilot discovery, who is permitted to create and share Copilot Studio agents, and whether Copilot access is explicitly scoped to managed devices only. These three decisions do not need to be perfect on day one, but they must be made. Deploying Copilot without answering all three is choosing to accept unknown data exposure at AI retrieval speed. That is a business decision, not an IT decision — it must be documented as a risk acceptance, not treated as a configuration gap to clean up later.
"The three controls that matter for Copilot governance aren't in the Copilot settings page. They're in SharePoint, Power Platform, and Entra."

🚫 What This Technology Does NOT Solve

Restricted Content Discovery does not fix overshared permissions. It hides content from Copilot's discovery surface, which buys time, but the underlying access model remains broken. A user can still navigate to the site, read the document, and manually paste it into any AI tool. The fix requires actual permission remediation, and that is a data governance project, not a Copilot feature toggle.

Conditional Access and tenant restrictions do not prevent a user from typing sensitive information into Copilot from a managed device. If a user pastes a confidential document into a Copilot prompt, the control boundary has already been bypassed. Microsoft Purview DLP policies for Copilot interactions are the complementary control for that scenario, and they deserve their own treatment.

Copilot Studio agent governance controls do not audit what existing agents are already doing. If agents were created before governance controls were implemented, those agents continue to operate. Organizations must run a discovery pass across Power Platform environments to identify agents already in production before tightening creation policies.

⚡ Assumption Challenge
Most organizations believe: "Once we configure these three controls, our Copilot governance posture is solid."
Reality: These three controls address the discovery surface, agent creation, and device boundary. They do not address data classification maturity, SharePoint permission remediation, Copilot interaction monitoring, or the gap between what Purview labels enforce and what Copilot can still retrieve. Governance is a system. Three controls are a starting point, not a destination.
ControlWhat It AddressesWhat It Does NOT Address
Restricted Content DiscoveryCopilot surfacing overshared sitesUnderlying permission sprawl
Copilot Studio agent governanceUnauthorized agent creation and sharingExisting agents already in production
Tenant Restrictions v2 + CACopilot on unmanaged devices or personal accountsUser-initiated data pasting within Copilot
Sensitivity Labels (Purview)DLP enforcement on labeled contentDiscovery of unlabeled content
🎯 Enterprise Decision Point
These controls are tactical mitigations, not a governance architecture. Organizations must decide whether they are accepting these three controls as permanent operating posture or as a bridge while they run a deeper SharePoint permissions remediation program. The answer determines the review cycle, the ownership model, and the audit scope. Running Restricted Content Discovery indefinitely on dozens of sites is a signal that permission remediation never happened — not a sign that governance is working.

🎯 Final Architect Recommendation

Deploy these controls now, before Copilot licensing goes broad. Not because the threat is theoretical, but because fixing governance gaps after users have already built workflows around Copilot responses is significantly harder than setting the posture at the start.

The sequencing I recommend: run the SAM data access governance report first. It will show you exactly which sites carry the most risk. Configure Restricted Content Discovery on the highest-risk sites immediately. Simultaneously, lock Copilot Studio agent sharing to require admin approval. Set the Conditional Access policy requiring managed device compliance for Copilot access before licenses are distributed to the first production wave.

The permission remediation underneath is a longer program. Restricted Content Discovery buys you the runway to do it properly without racing against a Copilot deployment timeline.

If I were advising an organization that had already deployed Copilot broadly without these controls, I would start with a Copilot interaction audit via Purview, identify what content has been surfaced in the last 90 days, and use that signal to prioritize which permissions to fix first. Reactive governance informed by real data beats a speculative permissions audit every time.

What I would not do: retrofit governance controls after the fact without executive visibility into the risk the organization has already accepted. The CISO needs to know the current exposure before she can sign off on the remediation plan. Keeping that conversation internal to IT is how organizations end up with an audit finding instead of a remediation story.

"Copilot governance isn't a Copilot project. It's a permissions project that Copilot finally made urgent."

🎯 The Takeaway

  • If Copilot is being deployed, run a SharePoint Advanced Management data access governance report first. If the report output surprises anyone in the room, Copilot governance is not ready.
  • Always configure Restricted Content Discovery on high-sensitivity sites before licenses reach end users. The alternative is discovering exposure through a prompt result, not a security review.
  • If Copilot Studio licensing is included in the rollout, restrict agent publishing to organization-wide audiences until a formal review and approval process is operational.
  • Always scope Conditional Access policies explicitly to include Copilot application endpoints. Do not assume Copilot inherits access controls from other Microsoft 365 policies without verifying the application scope.
  • If these three controls are implemented without a parallel permission remediation program, treat them as temporary mitigations with a defined review date, not permanent governance. Governance debt compounds the same way technical debt does.

Read more