Microsoft — Copilot summarized confidential emails despite DLP labels

AI relevance: A bug in Microsoft 365 Copilot Chat caused an AI assistant to process confidential, labeled emails, showing how LLM-driven agents can bypass enterprise data-loss policies when context boundaries fail.

  • Microsoft confirmed that Copilot Chat incorrectly summarized emails with confidential sensitivity labels applied, even when DLP policies were configured to block that access.
  • The issue affected the Copilot “work tab” experience, pulling content from Outlook Drafts and Sent Items folders.
  • Microsoft tracked the incident under service advisory CW1226324 and said it first detected the behavior on January 21.
  • A code issue allowed labeled emails to be processed, bypassing the intended label-based exclusion for Copilot chat.
  • Microsoft began a targeted fix rollout in early February and later stated the root cause had been addressed for most customers.
  • The company said access controls remained intact but acknowledged the behavior did not meet the intended Copilot experience for protected content.

Why it matters

  • Sensitivity labels are a core control for AI deployment in enterprises; bypasses turn AI features into unintended data access paths.
  • Copilot’s access to email content means a single misconfiguration or bug can expose highly sensitive material across workflows.
  • AI copilots embedded in office suites operate as privileged internal agents, so context-scoping failures become security incidents.

What to do

  • Review Copilot access: validate that sensitivity labels and DLP rules are enforced in Copilot chat and search experiences.
  • Audit high-risk folders: add monitoring for Drafts/Sent Items ingestion by AI features when confidentiality labels are set.
  • Implement compensating controls: restrict Copilot in regulated mailboxes until policy enforcement is verified.
  • Track advisories: ensure service messages like CW1226324 are fed into AI risk registers and incident reviews.

Sources