Copilot and overly broad sharing

What Microsoft 365 Copilot sees on the user's behalf, how overly broad sharing comes about, and what sensitivity labels add before the rollout.

Topic
AI
Author
ARLITECH expert team
Published

Microsoft 365 Copilot grants no one new permissions. It finds and summarizes what the user can already access: their mail, their chats, their OneDrive, and every SharePoint site and Teams team they were added to and never removed from. It is common for a user to have access to far more documents than they know about.

As long as a file can only be found by someone who knows which folder to look in, overly broad sharing rarely causes trouble. Copilot takes that obscurity away.

What Copilot sees on a user's behalf

Copilot reads through Microsoft Graph, always on behalf of the signed-in user. For any question it can draw on the sources the user could open with their own account:

  • their own mailbox and calendar, including the messages sent to them and by them,
  • the Teams chats and channels they are a member of, and meeting transcripts where one was recorded,
  • their own OneDrive and the files shared with them,
  • every SharePoint site they have at least read access to, whether directly or through a group membership.

The administrator decides whether Copilot may also use the public web for its answers. If this is allowed, the search queries Copilot writes are sent to Bing, which means they leave the Microsoft 365 service boundary. According to Microsoft's documentation, apart from this, prompts and responses stay within the service boundary and are not used to train the foundation models. All of this concerns what leaves the tenant. Who gets what from Copilot inside the tenant is decided by permissions.

An example: if a payroll spreadsheet sits on a site where the whole organization has been given read access, Copilot may draw on that file to answer “how much do team leads earn?” In that case Copilot is doing exactly what the permissions allow. This is why the principle of least privilege is the starting point for permission design.

How overly broad sharing comes about

Overly broad sharing is rarely deliberate. Unneeded access builds up over the years, from decisions made for convenience and cleanup that never happened. The typical sources:

The “Everyone except external users” group

This built-in group contains every internal user in the tenant. If it appears in the permissions of a site or folder, everyone in the organization can read the content. It easily happens that an old intranet page or a hastily set up project site is granted to this group and nobody takes it back later.

Public Microsoft 365 groups and Teams teams

Anyone in the organization can join a team set to public, and members of the organization can reach the content of its SharePoint site without joining. Public or private is settled in a drop-down when the team is created, and at that point the creator is not necessarily thinking about which files will end up there.

Organization-wide links

A “People in your organization” link opens the file for anyone in the company who uses the link. If such a link circulates in an all-staff email or a large Teams channel, the number of people with access grows without the file owner seeing it.

Broken inheritance and unique permissions

Within a site, individual folders and files can be given their own permissions. Over time these become impossible to follow: the site owner sees narrow rights at site level while a subfolder is shared with half the organization.

Accumulated rights and orphaned sites

An employee changes departments and joins the new groups but is never removed from the old ones. A project closes, the site stays, and its owner has left the company in the meantime. Nobody is then left to review the access, and the site's content remains available for years to everyone who was ever added.

Discovery: sites, groups, links

The goal of discovery is a list: which sites hold sensitive content, and who can access them. Microsoft 365 provides several tools for this. The data access governance reports in SharePoint Advanced Management show the sites that are shared with the “Everyone except external users” group, the sites where many organization-wide or anonymous links have been created, and those where the number of people with access is unusually high. The data risk assessments in Microsoft Purview Data Security Posture Management (DSPM) for AI look at overly broad sharing specifically from Copilot's point of view.

The report is only the starting point. For each flagged site, you need to settle the following:

  1. Does the site hold content that is not meant for everyone who can access it? Typically payroll and HR material, contracts, proposals, management reports, internal investigations, system documentation.
  2. Who actually owns the site and can make decisions about access? For an orphaned site, an owner has to be assigned first.
  3. Which access is still justified? A site access review lets you hand that decision to the site owner, who knows the content.

Inactive sites can go on a separate list: anything nobody has opened in years is best archived or deleted before it becomes a source for Copilot. For groups, discovery includes the list of public teams and a review of guest members. Guest account management is covered in more detail in the article on Microsoft 365 tenant security settings.

If the permissions on a sensitive site cannot be fixed quickly, two interim tools help. Restricted content discovery removes the site from Copilot and organization-wide search results, but it leaves existing access unchanged. The user's own and recently opened files can still show up, and Copilot still works with a file that is currently open. Restricted site access limits who can open the site at all to a specified group. Neither replaces cleaning up the permissions; they only buy time for it.

Sensitivity labels and encryption

Permissions control who can access a location. A sensitivity label marks the document or email itself and travels with it, even when the file is downloaded or forwarded. In Microsoft Purview, a label can mean several things:

  • marking: a header, footer, or watermark that is visible on the document,
  • encryption with usage rights: the label defines who can open the content and what they can do with it (view, edit, print, copy content),
  • a condition for other policies, for example data loss prevention (DLP) policies.

A label without encryption does not restrict access on its own. If the “Confidential” label only adds a watermark to the file, Copilot still uses it for any user who has access. With an encrypting label, Copilot can only work with the content if the user has the right to copy content as well as to view it. For the most sensitive categories, the encrypting label is therefore what sets a real limit.

Two more behaviors belong here. When Copilot creates new content from labeled documents, for example from files referenced in Word or PowerPoint, the new document inherits the highest-priority source label in supported scenarios. A DLP policy set up with Microsoft 365 Copilot as its location can stop Copilot from using content with specified labels in the generated response. The file can still appear as a reference, and the policy does not cover every Copilot experience.

Labels are only worth something if documents actually get them. A rollout usually starts with a small set of labels (for example Public, Internal, Confidential, Highly confidential) and a default label at site or library level, and it tries auto-labeling in simulation mode first to find out how many documents it would label incorrectly.

Pilot group and phased rollout

Copilot is turned on per user, through licenses, so the rollout is easy to phase. For the pilot group, pick employees from several departments who will report a problem instead of working around it. Include people whose work involves searching through a lot of documents, because they are the first to notice when Copilot cites a source they should not have access to.

From the very start of the pilot you need a simple reporting channel for “odd Copilot results”, for example a dedicated Teams channel or a helpdesk category. Some of the sites reported this way will be ones that did not stand out in the reports. During the pilot, keep an eye on:

  • user reports that Copilot surfaced an unexpected document,
  • the unified audit log, where Copilot interactions and the referenced sources also appear,
  • the SharePoint Advanced Management reports and the DSPM for AI assessments, because fixing the problem sites found during the pilot is a precondition for the next round.

The expansion goes department by department, and before each round that department's sites are checked. Payroll and HR, legal, and management sites should come first: protect them with restricted content discovery or an encrypting label before Copilot is turned on for a larger part of the organization. End each round with a decision point where the list of open issues determines whether the next group can go ahead.

What users need to know

Using Copilot does not take long training, but there are a few things every user needs to know on day one:

  • Copilot only sees what you can access yourself. If it shows you a document you should not be able to see, report it through the reporting channel, and do not pass the content on or use it.
  • Answers have to be checked. The language model sounds convincing even when it is wrong, so open the cited sources before a figure or a claim goes into an email or a decision.
  • What Copilot creates is your document, and how it is shared and labeled is your responsibility.
  • When you share your own file, pick the narrowest link: specific people, not the entire organization.
  • The corporate Copilot and public chatbots used with a personal account are not the same thing. Company data may only be entered into the approved tool. The article on shadow AI covers this.

Related service: Secure Copilot and Microsoft 365 AI rollout →

Contact

Ready to secure your critical systems?

Our experts will assess where your security stands and build a roadmap that fits your organization.

ARLITECH's office in Gödöllő

Visit us

You are welcome to visit our Gödöllő office in person.

Address
Centrum Office Building, 2100 Gödöllő, Dózsa György út 13, 3rd floor, office 309
Office hours
Mon–Fri, 9 a.m. to 5 p.m.
Get directions (opens in a new tab)