Why does Microsoft 365 oversharing matter for Copilot?
Microsoft 365 Copilot and Copilot Studio agents answer with content the signed-in user can already open. A file shared through an 'Anyone' link, an organization-wide link, or a forgotten guest grant can surface in a Copilot answer to people who were never meant to see it. Scoring every SharePoint site and OneDrive for that exposure, and revoking the riskiest links first, is how you get the data ready for Copilot.
What signals go into a site's exposure score?
Five weighted factors: Exposure (anonymous and organization-wide links, up to 40 points), External reach (guests and outside recipients across domains, up to 20), Blast radius (storage held and the groups granted access, including role-assignable and dynamic groups, up to 20), Dormancy (sharing left on sites nobody is using, up to 15), and Hygiene (links with no expiry or direct grants add points; download-blocked links take them away, -3 to +9). The factor breakdown is always shown, so the number is never a black box.
What do the risk levels mean?
Critical is above 70, High is 51-70, Medium is 31-50, and Low is 0-30. A site whose permission scan is incomplete is marked Insufficient data rather than Low, because an unread file can hold links the score has not seen.
Does scoring or remediation change anything without approval?
No. Scoring is entirely read-only. Revoking a link or running a bulk revoke is an explicit, admin-initiated action - Sentinel never removes access on its own. Bulk revoke always leaves owners their access, and every action is written to the Revoke Activity log.
How fresh are the scores?
Scores are recalculated on a schedule you control, against the inventory the sync has pulled from Microsoft Graph; the sync reads only what changed since the last run. After you remediate, the affected sites are re-scored on the next run so the queue reflects reality.
What are the AI summaries, and where does that run?
On top of the deterministic score, an optional Azure OpenAI deployment in your own subscription writes a plain-language summary of the tenant and, on request, of a single site. Summaries are stored and reused until the score behind them changes. The score and the action items never depend on AI, and you can deploy without Azure OpenAI at all.