A tenant-wide consent to an AI vendor is a legitimate administrative action, so no single log reads its exercise as exposure. The directory logs the grant and the shared-account sign-in without naming the human behind the token. The directory activity shows the grant in use, 3,180 requests opening with full enumeration, without weighing what was pulled. The content audit names the files read while the Vendor Risk Catalog resolves the counterparty, a 19-day-old vendor with no verified publisher, no SOC 2, and no data-processing agreement. Triangulated, the sources establish that a permanent grant to an unassessed vendor is running through an account no human owns, with the catalog setting severity and the audit record setting scope.
.avif)
