Microsoft confirmed that a bug allowed Copilot Chat to read and summarize customers' confidential emails for weeks, even when data-loss prevention policies were supposed to block the processing of sensitive material. The affected messages included drafts and sent emails marked confidential. Microsoft began rolling out a fix in February and did not say how many customers were affected, TechCrunch reported on February 18th.

The office assistant was given a rule and then crossed it without announcing the breach. That detail changes the usual picture of AI as a voluntary helper waiting for a question. Copilot was embedded in the software where the correspondence already lived, and a failure in the surrounding permissions turned private writing into material available for summary. A confidentiality label can look like a promise to the person who applies it and a field in a database to the system processing it. The bug joins a growing list of reasons institutions are blocking built-in AI on work devices. The envelope remains closed on the desk, but its contents have already become an answer inside another window.