Copilot bug summarised confidential emails despite DLP
A code error let Microsoft 365 Copilot Chat summarise emails labelled confidential, despite DLP rules. Why a label should not be the only safeguard.
Published · 3 min read

On 18 February 2026, BleepingComputer reported a Microsoft service advisory, CW1226324, saying that Microsoft 365 Copilot Chat had been summarising emails that carried a confidential label, even where a data loss prevention (DLP) policy was configured. A code issue let items in users' Sent Items and Drafts folders be picked up by Copilot despite their labels. The problem was first detected on 21 January, and Microsoft began rolling out a fix in early February, without saying how many organisations were affected. For an office that relies on labels to keep sensitive mail away from an AI feature, it shows where that protection lives: in the feature's own code.
Microsoft said on 19 February that the emails involved were the user's own, written by them and stored in their Drafts and Sent Items, and that the bug gave nobody access to information they were not already allowed to see. It also said the behaviour fell short of what it intended for Copilot, which is meant to keep protected content out of its reach. On 20 February it said it believed the root cause had been addressed for most customers, with deployment continuing in a small section of its more complex environments.
A label is only as good as the code that reads it
The labels on those emails were designed to restrict access by automated tools, and the DLP policy was the rule telling Copilot to leave them alone. A rule like that is carried out by the feature's own code. From at least 21 January until the fix reached each organisation, that code had a gap in two folders, while the labels stayed where they were.
Copilot Chat began reaching Word, Excel, PowerPoint, Outlook and OneNote for paying Microsoft 365 business customers in September 2025, BleepingComputer notes, so for offices on those plans the AI arrived inside software they already paid for. That makes switching it on, and deciding what it may read, a decision to take on purpose. Under the Data Privacy Act of 2012 (Republic Act 10173), retrieving, consulting and using personal information all count as processing, and Section 20 asks for reasonable and appropriate measures to protect it. A label is one such measure. It should not be the only one.
What it means for the AI features I build
Storia, the wedding-planning platform I run, has an assistant that can change a couple's budget, wedding date, timeline and guest list, and it saves none of it on its own: every change is shown to the couple first and saved only when they confirm. Some data never reaches the AI at all. On a lawyer's advice, guests' dietary needs are never sent to it, and a guest's details reach it only when the server finds a consent record. If the record is missing, the details are withheld.
The design point is where each limit sits. A rule an AI feature is asked to follow can fail the way any code can, and data it is never given cannot reach it through a bug. The same two questions apply to any AI feature an office buys or builds: what is kept out of its reach entirely, and does it change records itself or only propose changes for a person to confirm?
Five questions before an AI feature is switched on
- What can it read? List the mailboxes, sites and folders it actually reaches, which may be more than the ones you had in mind.
- What must it never read? Keep HR cases, health information and similar material where the feature has no access at all, rather than behind a label alone.
- Can you see what it read? Find out whether its use is logged, and who reviews the log.
- Does it act, or propose? A feature that drafts for a person to check carries less risk than one that sends, saves or deletes.
- Who decided, and for whom? Put the decision to switch it on in writing, with the groups it covers, and add it to your Data Privacy Act inventory.
If you are deciding whether to switch one on, the fit check is where I would begin.
Sources
- BleepingComputer, Microsoft says bug causes Copilot to summarize confidential emails, 18 February 2026
- Petri, Microsoft 365 Copilot Bug Exposes Confidential Emails Despite DLP Safeguards, 19 February 2026
- Lawphil, Republic Act No. 10173
Enrique Lacambra III builds internal business systems from Cainta, Rizal.
Tell me what is broken and I will tell you if it is worth building. A 20-minute call, no charge, and I will say so if you do not need me.
