For a period of four weeks commencing on January 21st, Microsoft’s Copilot had the ability to access and summarize confidential emails, despite all sensitivity labels and DLP policies specifically prohibiting it from doing so. This breach occurred within Microsoft’s own systems, with no security tool detecting the unauthorized access. This incident impacted several organizations, including the U.K.’s National Health Service, highlighting the severity of the breach in regulated healthcare environments. Known internally as CW1226324, this breach was only discovered and reported in an advisory by BleepingComputer on February 18th.
This is not the first time Copilot has breached its trust boundary. In June 2025, Microsoft patched a critical vulnerability dubbed “EchoLeak,” which allowed malicious emails to exfiltrate enterprise data without any user interaction. The root causes of these breaches are a code error and a sophisticated exploit chain, both resulting in Copilot accessing restricted data.
The lack of detection by traditional security tools such as EDR and WAF is due to the fact that they were not designed to monitor AI assistants like Copilot. These tools monitor file behavior and inspect HTTP payloads, but they do not have the capability to detect violations of trust boundaries by AI systems.
To address these vulnerabilities, organizations are advised to conduct regular tests of DLP enforcement against Copilot, block external content from reaching Copilot’s context window, audit Purview logs for anomalous interactions, enable Restricted Content Discovery for sensitive data, and establish an incident response playbook for vendor-hosted inference failures.
It is crucial for organizations to be proactive in addressing these vulnerabilities, as AI assistants continue to be deployed at a rapid pace without adequate governance. By implementing the recommended controls and audit measures, organizations can better protect sensitive data and mitigate the risk of AI trust boundary violations.
