Get to know our comprehensive Cybersecurity Portfolio: Learn More

close icon

Conozca nuestro completo portafolio de ciberseguridad: Aprenda más

How to Investigate Data Exposure Through Microsoft Copilot o AI Agent Identity Security with Microsoft Entra ID.

Toggle

To investigate data exposure through Microsoft Copilot, locate the interaction in the unified audit log, identify the source documents Copilot referenced, verify how the user had access to them, and determine who else shares that access. Then assess the impact on personal and regulated data and fix the root cause in permissions, sensitivity labels and DLP policies.

Talk to our experts in Secure Enterprise AI for Microsoft Environments

The scenario usually starts with a message to IT or HR. An employee asks Microsoft 365 Copilot to “summarize recent compensation discussions” and receives a response that includes salary figures from a payroll workbook stored on a SharePoint site. The employee was never supposed to see that data. What happens next determines whether the organization contains a data risk calmly, or disables Copilot for everyone and never finds the real cause.

What Is Data Exposure Through Microsoft Copilot?

Data exposure through Microsoft Copilot occurs when Copilot surfaces sensitive data, such as financial data, PII, PHI, legal documents or HR records, to a user who should not have it. It happens even though that user technically has access to the underlying content.

Copilot uses large language models grounded in organizational content retrieved through Microsoft Graph. The LLM does not decide who can see what. The permissions on emails, files, sites and chats do.

Does Copilot Bypass Permissions?

No. Microsoft 365 Copilot only retrieves content the user is already authorized to access. The problem is that many organizations have years of oversharing that nobody noticed:

  • Sites shared with “Everyone except external users”
  • Broken permission inheritance
  • “Anyone with the link” sharing
  • Sensitive files with no sensitivity labels

Before Copilot, that content was technically accessible but practically hidden, because nobody knew where to look. Copilot removes that friction. One natural-language question can surface what a user could have found manually but never would have. Copilot data exposure is therefore usually the visible symptom of an underlying permissions and data governance problem.

Common Causes of Copilot Data Exposure

Cause Example Detection source Remediation
Broad site access Finance site shared with “Everyone except external users” SharePoint data access governance reports, DSPM data risk assessments Restrict membership; apply least-privilege access
Sharing links Payroll file shared via an organization-wide link Sharing reports, audit log sharing events Remove or scope links; set link expiration
Broken inheritance Folder with unique permissions inside a restricted library Permission reports Restore inheritance or reassign permissions
Missing labels HR files with no sensitivity label or encryption Activity Explorer, data classification reports Apply labels with encryption; enable auto-labeling
Agents with excessive grounding Copilot Studio agent grounded on an entire HR site Agent inventory, Copilot audit records Narrow knowledge sources; review agent permissions

Evidence Sources for Investigating Copilot Exposure

A Copilot investigation draws on several sources in Microsoft Purview and Microsoft 365.

Unified audit log. Copilot activity is recorded automatically as part of Audit (Standard) under these record types:

  • CopilotInteraction: Microsoft Copilot experiences, including Copilot Studio agents.
  • TeamCopilotInteraction: Teams facilitator features.
  • ConnectedAIAppInteraction: custom or third-party AI apps registered in your organization.
  • AIAppInteraction: external AI tools, logged under pay-as-you-go billing.

Each record identifies the user, the time, the app where the interaction occurred and the resources Copilot accessed to generate the response, including sensitivity label IDs where they exist. It also shows whether Copilot used web search.

Prompt and response content. The audit record shows that an interaction happened and what Copilot referenced. The prompt and response text is retrieved through eDiscovery and, where collection is configured, through DSPM and Activity Explorer.

Microsoft Purview DSPM. Data Security Posture Management provides AI activity insights and data risk assessments that identify overshared sites and unlabeled sensitive content. Microsoft has announced the retirement of the classic DSPM and DSPM for AI experiences (message center post MC1481315). The most recent update moves the retirement to January 2027, so investigation workflows should be built in the unified DSPM experience.

Activity Explorer. Shows labeling activity, sensitive information type matches and AI interactions involving sensitive data.

eDiscovery. Preserves and exports Copilot prompts and responses, and supports legal holds on the users involved.

SharePoint permission and sharing reports. Data access governance reports in SharePoint Advanced Management show oversharing at site level.

Defender XDR advanced hunting. Useful when the investigation extends to agent activity or a compromised account.

How to Investigate a Copilot Data Exposure in 7 Steps

Step 1: Capture the Report and Preserve Evidence

Record who reported the exposure, when, the approximate prompt and what the response contained. Ask the user not to delete the conversation. If the exposure may involve personal data, litigation or regulatory reporting, place the user’s mailbox (where Copilot interaction data is stored) and the relevant sites on legal hold through eDiscovery.

Copilot audit records are not covered by the one-year default retention policy of Audit (Premium). They are retained for 180 days unless a custom retention policy extends them, so preserve them early.

Step 2: Locate the Copilot Interaction in the Audit Logs

In the Microsoft Purview portal, open Audit and filter by user, date range and the CopilotInteraction record type. The AppHost field shows where the interaction happened, such as Microsoft 365 Copilot Chat, Teams, Word or Outlook. Export the matching records for analysis.

Step 3: Identify the Source Documents Referenced in the Response

Review the accessed resources in the record’s CopilotEventData. These references list the files, emails, sites or chats Copilot used to generate the response, often with file names, locations and sensitivity labels. In the payroll scenario, this would show the exact workbook and SharePoint site Copilot drew from. Retrieve the response text through eDiscovery to confirm which information was actually disclosed.

Step 4: Verify How the User Had Access

This is the core of the investigation. For each source document, determine the access path:

  • Direct permission
  • Group membership
  • An organization-wide sharing link
  • Inheritance from a broadly shared site
  • A broken inheritance chain

The answer separates two very different outcomes. Either the user had legitimate access (no exposure, but possibly a policy question) or the access exists because of oversharing.

Step 5: Determine the Exposure Scope

If one user could reach the file, others probably could too. Identify everyone who has the same access path, and use the audit log to check whether anyone else accessed the file directly or referenced it through Copilot. This step defines the blast radius: the gap between who should have access and who actually does.

Step 6: Assess Impact

Classify what was exposed: PII, PHI, financial data, legal privileged material, trade secrets. Note whether the content was labeled or protected with encryption. Then determine how many people saw it, or could have, and whether the information left the organization. This assessment drives the regulatory analysis described below.

Step 7: Remediate the Root Cause and Contain

Fix the source rather than the symptom:

  • Remove broad access and expired sharing links
  • Restore permission inheritance
  • Apply sensitivity labels with encryption
  • Extend DLP policies to Copilot

For sensitive sites that cannot be fixed immediately, restricted content discovery prevents them from appearing in Copilot results, and restricted access control limits site access to defined groups. DLP for Microsoft 365 Copilot can also stop Copilot from processing labeled content or responding to prompts that contain sensitive information types, which makes it useful as a short-term containment control.

Investigating Data Exposure Through Copilot Studio Agents

Agents add a layer of complexity. A Copilot Studio agent may be grounded on SharePoint sites, connectors or external data. Depending on how it is configured, it may answer using the agent’s own access instead of the end user’s permissions.

When an agent is involved, investigators should establish:

  • Which knowledge sources and connectors the agent uses
  • Which authentication mode it uses
  • Who published it
  • Who can interact with it

Copilot Studio interactions appear in the audit log under CopilotInteraction, with an app identity that names the agent. Microsoft Entra Agent ID is designed to give agents their own identities, which makes it possible to trace agent activity and govern agent permissions in Entra just as you would for users and applications.

An agent grounded on an entire HR site and published to the whole organization effectively republishes that site to everyone who can chat with it. Agent configuration reviews should be part of every exposure investigation.

Is a Copilot Data Exposure a Reportable Data Breach?

It can be. Under GDPR, a personal data breach includes unauthorized disclosure of or access to personal data, and internal disclosure to an unauthorized employee can qualify. If the breach is likely to result in a risk to individuals, the controller must notify the supervisory authority within 72 hours of becoming aware of it. All breaches must be documented, even when notification is not required.

Under HIPAA, unauthorized access to PHI by a workforce member is presumed to be a breach unless a documented risk assessment shows a low probability that the information has been compromised. Some narrow exceptions apply.

Whether a specific Copilot exposure is reportable depends on:

  • What data was exposed
  • Who saw it
  • Whether it was further used or disclosed
  • The applicable regulations and contracts

Involve your data protection officer and legal counsel early. The investigation steps above produce exactly the facts they need, and doing them quickly protects the notification timeline.

How to Prevent Future Copilot Data Exposure

Investigations should feed a prevention program, not end with a single fix:

  • Run oversharing assessments regularly using DSPM data risk assessments and SharePoint data access governance reports, starting with sites that hold financial data, HR records and PHI.
  • Strengthen information protection with sensitivity labels, encryption and auto-labeling based on data classification.
  • Extend data loss prevention to Copilot for labeled content and sensitive prompts.
  • Use restricted content discovery for high-risk sites while remediation is underway.
  • Monitor risky AI usage with Insider Risk Management, which can detect patterns such as repeated prompts seeking sensitive information.
  • Govern agents with approved knowledge sources, publishing controls and periodic reviews.
  • Address AI-specific threats such as prompt injection, which can manipulate Copilot into retrieving or leaking content.

The broader technical controls are covered in ne Digital’s Microsoft 365 security hardening guide. The AI-specific risks are analyzed in Is Your Microsoft Copilot Deployment Vulnerable?. Over the long term, these controls belong in an AI governance framework aligned with the NIST AI RMF and ISO/IEC 42001. For organizations using more than one AI platform, ne Digital explains how to govern Claude, ChatGPT and Microsoft Copilot under one AI framework.

Conclusion

Copilot data exposure is rarely a Copilot problem. It is the moment an organization’s existing oversharing becomes visible. The right response is a structured investigation:

  1. Preserve the evidence.
  2. Reconstruct the interaction.
  3. Trace each source document to its access path.
  4. Measure the blast radius.
  5. Assess regulatory impact.
  6. Remediate at the source.

Disabling Copilot hides the symptom while leaving the same sensitive data accessible to the same people.

Talk to our experts in Secure Enterprise AI for Microsoft Environments

ne Digital helps organizations investigate AI-related data exposure, remediate oversharing with Microsoft Purview and SharePoint controls, and build AI governance frameworks aligned with the NIST AI RMF and ISO/IEC 42001. A Copilot data exposure assessment is the most direct way to find and close these gaps before the next prompt does it for you.

Topics: Artificial Intelligence

Frequently Asked Questions about Data Exposure Through Microsoft Copilot o AI Agent Identity Security with Microsoft Entra ID

Can Microsoft 365 Copilot expose confidential data?

Yes, when that data is overshared. Copilot surfaces content the user can already access, so broadly shared sites, sharing links or broken inheritance can lead to confidential information appearing in responses.

How do I see what a user asked Copilot and what it answered?

Use the CopilotInteraction records in the unified audit log to identify the interaction and the resources referenced. Use eDiscovery to retrieve the prompt and response content.

How do I find which documents Copilot used in a response?

Review the accessed resources listed in the Copilot audit record. They identify the files, sites and messages Copilot referenced, including sensitivity labels where applied.

What is Microsoft Purview DSPM?

Data Security Posture Management in Microsoft Purview provides visibility into sensitive data and AI activity, including data risk assessments that identify overshared content. The classic DSPM for AI experience is being retired in favor of the unified DSPM.

Should we disable Copilot after a data exposure incident?

Usually not. Disabling Copilot does not remove the underlying oversharing. Targeted containment, such as restricted content discovery, DLP for Copilot and permission fixes, addresses the real cause.

Can Copilot Studio agents expose data to unauthorized users?

Yes, if they are grounded on broad knowledge sources or published widely. Review each agent’s knowledge sources, authentication mode and audience as part of the investigation.

How long are Copilot interaction logs retained?

Copilot audit records are retained for 180 days by default. Longer retention requires a custom audit retention policy with Audit (Premium). ne Digital can help configure retention so that evidence is available when an investigation starts.

Related Articles

Based on this article, the following topics could spark your interest!

AI Strategy Roadmap: How Enterprises Tur...

The biggest mistake organizations make with AI is not techni...

Read More
Types of Artificial Intelligence: Do We ...

In 2026, the phrase Types of Artificial Intelligence is used...

Read More
The Role of Artificial Intelligence in C...

With the rise of cyber threats, the need for robust security...

Read More