3.3. Protecting Sensitive Data with Copilot
💡 First Principle: Copilot can only expose what you already have permission to see — but it can surface that content more efficiently and more broadly than traditional search. The risk is not that Copilot breaks security; it is that it removes friction that previously made over-sharing harder.
This is a nuanced but important distinction. Microsoft 365 Copilot inherits your existing permission model — it does not bypass it. If you cannot access a file, Copilot cannot show you that file. However, if you have broad access to sensitive information (perhaps more than you typically use), Copilot can surface it in ways that feel surprising.
Common sensitive data risks with Copilot:
| Risk Pattern | Example | Mitigation |
|---|---|---|
| Over-broad permissions | You have read access to all of HR's SharePoint, so Copilot includes salary data in a summary | Apply least-privilege permissions; sensitive files should be restricted |
| Unintentional disclosure in shared output | You share a Copilot-generated summary that includes content from a restricted document | Review AI output for sensitive content before sharing |
| Prompt context exposure | You paste sensitive data into a prompt to get help analyzing it, and that prompt history is accessible | Understand your organization's Copilot history and retention settings |
| Agent knowledge misconfiguration | An agent is configured with a knowledge source that includes sensitive files accessible to too many users | Carefully scope agent knowledge to appropriate audiences |
The principle of least privilege applies directly to Copilot usage: the less sensitive content a user can access, the less sensitive content Copilot can surface in their experience. Good data governance and permission hygiene upstream of Copilot reduces risk downstream.
⚠️ Exam Trap: Data protection policies restrict what Copilot outputs, not just what it accesses. Microsoft Purview sensitivity labels and DLP (Data Loss Prevention) policies can prevent Copilot from generating responses that would expose protected content — even to users who technically have permission to access the underlying files. This is a two-layer protection: access control AND output filtering.
Reflection Question: A team member reports that Copilot surfaced salary information in a response to a general question about the HR team. What is the most likely cause, and what organizational change would prevent this?
