Risk type: Tool Risk
Microsoft Copilot security
Microsoft 365 Copilot only surfaces data users can view. Oversharing becomes Copilot exposure. Prompt injection is the second risk.
Microsoft says prompts do not train its foundation models. Copilot honors SharePoint permissions and label encryption. So fix broad sharing before rollout. CVE-2025-32711 let crafted content expose Copilot data. Microsoft mitigated it, and users had nothing to do. Review sharing, labels, and allowed agents first.
Last updated
2026-10-30
Copilot risks and the first control for each
| Risk | Where it shows up | First control |
|---|---|---|
| Oversharing | Sites shared with everyone surface in answers. | Run data access governance reports, starting with Everyone except external users. |
| High-risk sites | Sensitive sites appear in Copilot answers. | Use restricted content discovery to keep them out of Copilot. |
| Unlabeled confidential files | Files lack encryption labels. | Apply sensitivity labels with encryption. Copilot honors their usage rights. |
| Prompt injection | Email or files carry crafted instructions, as in EchoLeak. | Limit agents and treat external content as untrusted. |
| Agents and connectors | Agents query data the user can access. | Allow only reviewed agents in the Microsoft 365 admin center. |
From Microsoft Learn and the Microsoft Security Response Center. Sources checked October 30, 2026.
Affected tools
- Microsoft 365 Copilot
- Copilot Chat
- Copilot agents and Graph connectors
- Copilot in Outlook and Teams
How it usually happens in the browser
- A user asks Copilot a broad question across the tenant.
- Copilot finds a file shared with everyone years ago.
- The answer quotes salary, legal, or deal data.
- Crafted content in email or files reaches Copilot.
- Copilot reads it while answering an unrelated question.
- Users click links in answers without checking them.
What traditional defenses miss
- Old sharing links look harmless until Copilot finds them.
- Permission reviews lag behind how fast sites get shared.
- AI answers feel authoritative, so people skip review.
- Prompt text can carry instructions that filters miss.
Mitigation checklist
- Run the oversharing reports in SharePoint Advanced Management.
- Fix “Everyone except external users” access on sensitive sites.
- Turn on restricted content discovery for high-risk sites.
- Apply sensitivity labels with encryption to confidential files.
- Allow only reviewed agents in the admin center.
- Review AI answers before they reach customers.
Where browser isolation fits
- Shield runs the page in a browser that is not on your computer, destroyed when you close the tab.
- Open untrusted links from Copilot answers in Shield.
- Isolation does not change what Copilot can retrieve.
- Microsoft 365 permissions decide what Copilot can see.
Limits
- Your licenses decide which SharePoint tools you can use.
- Microsoft’s injection classifiers do not cover every Copilot scenario.
- EchoLeak is fixed, but new injection flaws remain possible.
- We reviewed Microsoft documentation and did not test Copilot.
FAQs.
Does Microsoft 365 Copilot respect existing permissions?
Was EchoLeak fixed?
Does Copilot train on our prompts?
How do we reduce oversharing before rolling out Copilot?
Do we need to block Copilot to be safe?
References
- 01Data, Privacy, and Security for Microsoft CopilotMicrosoft Learn
- 02
- 03Secure & Governed Data Foundation for Microsoft CopilotMicrosoft Learn
- 04CVE-2025-32711: M365 Copilot Information Disclosure VulnerabilityMicrosoft Security Response Center
- 05OWASP Top 10 for Agentic Applications (December 9, 2025)OWASP Gen AI Security Project
Keep exploring
- AI data exfiltration through plugins and toolsChannels that move data out, including EchoLeak.
- Prompt injection: direct and indirect attacksHow hidden instructions steer a model.
- Sensitive data in AI promptsWhat each AI vendor keeps from your prompts.
- Install the skillAgent builders reach the same routing and isolation through a ready skill.
- All guidesThreats, secure apps, and AI security pages.
- All AI security guidesBrowse AI topics by risk type.
- Secure apps directorySecure browsing guides for top SaaS apps.
- Threat libraryThreat pages with checklists and FAQs.