Microsoft introduces new security policies for Copilot data grounding

Microsoft is implementing new security measures within Microsoft Purview Data Loss Prevention (DLP) to manage how Microsoft 365 Copilot accesses and utilizes organizational data. A key update allows organizations to restrict Copilot to grounding its responses using only trusted internal emails, rather than external messages.

This policy aims to mitigate risks such as cross-prompt injection, where external senders include hidden instructions designed to manipulate AI outputs. By excluding unverified external emails from the grounding process, the system reduces the likelihood of Copilot summarizing misleading content or following malicious instructions disguised as routine communication.

Technically, Copilot does not read files directly. Instead, it uses Microsoft Graph and Work IQ to ground prompts, ensuring that the AI respects existing user permissions, Restricted SharePoint Search settings, and Purview data labels. When web access is required, Copilot sends anonymized queries to Bing to maintain the privacy of the user's identity and specific work data.