Microsoft says Office bug shared confidential emails with Copilot AI

Microsoft has confirmed that a bug allowed its Copilot AI to access and summarize confidential customer emails without permission for several weeks.

The issue, first reported by Bleeping Computer, affected Microsoft 365 users since January, even if data loss prevention policies were in place.

The bug, identified as CW1226324, caused emails labeled as confidential to be improperly processed by Copilot Chat, which is available to paid Microsoft 365 subscribers for use in Word, Excel, and PowerPoint.

Microsoft said it began deploying a fix in early February, but did not specify how many customers were impacted.

🔗 Source: TechCrunch

🧠 Food for thought

Implications, context, and why it matters.

The flaw slipped past a data protection layer many companies count on

Microsoft’s security approach for Microsoft 365 Copilot depends on existing user access rules plus admin settings through tools such as Microsoft Purview Data Loss Prevention (DLP), which helps stop sensitive information from being shared or used inappropriately 1.
Microsoft framed the issue as a breakdown in the DLP enforcement layer, which should block Copilot Chat from processing content marked with sensitivity labels such as “Confidential” 2.
The vulnerability let Copilot Chat pull restricted emails from users’ “Sent Items” and “Drafts” folders even when those confidential labels were applied 2.

This case turns AI security concerns into a real-world problem

The behavior matches worries that have pushed some groups to limit Copilot, including the European Parliament’s IT department, which reported blocking AI features on government devices over concerns about possible cloud data uploads.
The episode reinforces the shared responsibility model for Microsoft 365 Copilot security, where Microsoft ships controls while customers must set governance such as permissions, labels, and DLP, then monitor their environments 3.
Generative AI can magnify a single policy failure because the tool quickly searches and combines information, which can turn an enforcement slip into active data exposure 2.
The outcome may raise the bar for enterprise reviews, push more buying of third-party security tools, and invite tougher regulatory scrutiny of AI providers’ data handling claims 4.