Did Microsoft Copilot Ignore Sensitivity Labels? A Deep Dive
Microsoft Copilot has ignored sensitivity labels twice in eight months, raising serious concerns about data security and trust in AI systems. This issue is particularly alarming in highly regulated environments, such as the U.K.'s National Health Service. For four weeks, starting January 21, Copilot read and summarized confidential emails, despite robust sensitivity labels and Data Loss Prevention (DLP) policies designed to protect sensitive information. Microsoft's own pipeline failed to flag these violations, highlighting a critical security gap.
What Incidents Occurred?
The first incident, tracked as CW1226324, involved Microsoft’s Copilot processing sensitive email content that it was instructed to skip. This was not an isolated failure; it marked the second time in eight months that Copilot's retrieval pipeline violated its own trust boundaries. The earlier incident was even more severe. In June 2025, Microsoft patched a critical vulnerability known as CVE-2025-32711, or "EchoLeak." This flaw allowed a malicious email to bypass multiple security layers and exfiltrate enterprise data without any user interaction.
Why Did These Failures Happen?
Both incidents resulted from a combination of a code error and a sophisticated exploit chain, leading to unauthorized access to restricted data. The security stack failed to detect these breaches because existing tools like Endpoint Detection and Response (EDR) and Web Application Firewalls (WAF) are not designed to monitor AI assistant interactions. This blind spot reveals significant weaknesses in the security architecture surrounding AI systems.
How Can Organizations Prevent Future Breaches?
To prevent similar breaches, organizations must implement a comprehensive five-point audit to ensure AI systems like Copilot adhere to security protocols. Here are actionable steps:
- Test DLP Enforcement Against Copilot: Regularly verify if Copilot honors sensitivity labels, especially on Sent Items and Drafts. Conduct these tests monthly to ensure compliance.
- Block External Content: Disable external email context in Copilot settings to prevent malicious content from reaching the AI assistant. This reduces the risk of prompt-injection attacks.
- Audit Purview Logs: Review Copilot interactions for unauthorized access during known exposure windows. This documentation is crucial for compliance and audit purposes.
- Enable Restricted Content Discovery: Use Restricted Content Discovery for SharePoint sites housing sensitive data to eliminate the risk of data entering Copilot’s context.
- Develop an Incident Response Plan: Create a playbook for incidents involving trust boundary violations within vendor-hosted inference pipelines. Assign ownership and establish monitoring protocols.
How Do These Failures Affect Businesses?
The implications of these failures are significant. Organizations relying on AI tools must recognize that their data security is only as strong as the tools they use. With 47% of senior security leaders reporting unauthorized AI behavior, governance must evolve alongside AI technology. The inability to monitor AI interactions effectively can lead to severe breaches of confidentiality, especially in regulated industries like healthcare.
What’s Next for AI Governance?
As businesses increasingly deploy AI assistants, they must prioritize security frameworks that address these new risks. The structure of AI systems typically includes a retrieval layer, an enforcement layer, and a generation layer. If any enforcement fails, sensitive data can be exposed without detection. Organizations must ensure robust testing and monitoring of their AI tools to prevent future breaches.
Conclusion: How Can Organizations Safeguard Their Data?
The recent failures of Microsoft Copilot serve as a wake-up call for organizations using AI technologies. Ensuring data security in AI systems is no longer optional; it’s critical. By implementing the five-point audit and redefining the security landscape, businesses can better protect their sensitive information against potential breaches.
In a world where data breaches can lead to severe financial and reputational damage, safeguarding against these vulnerabilities is paramount. Be proactive and take the necessary steps to ensure your organization’s data remains secure.
