How do you inventory unsanctioned AI browser extensions?

Pull the inventory from browser enterprise management rather than the endpoint. Chrome Enterprise Core reports installed extensions across managed profiles, and Microsoft Edge for Business provides the equivalent. Cross-reference that list against OAuth grants in your identity provider and against endpoint file inspection for unmanaged installs. Extensions are the highest-permission, lowest-visibility AI surface in most enterprises.

The permission model is what makes this urgent. A typical AI assistant extension requests read and change access to data on all websites, which grants it the ability to read every page a user visits, including your CRM, your ticketing system, your email, and your internal applications, and to transmit that content to a third-party endpoint. That is broader access than most sanctioned integrations hold, granted by a single click and no review.

Record four fields per extension: publisher and store listing, requested permissions, install count across your fleet, and whether it is present in your allowlist. Then check the ones nobody flagged, because AI functionality is now bundled into extensions that present as writing tools, meeting notetakers, tab managers, and shopping assistants.

Watch for post-install updates. Extensions update automatically, and permission scope can expand in a version the user never reviewed, which is the same rug-pull pattern that affects MCP servers.

Move to a force-install allowlist once you have the inventory. Blocklisting individual extensions is unwinnable because the store adds new ones faster than you can review them.

Akto Atlas discovers AI browser extensions alongside the rest of the employee AI estate, mapping each to the data it can reach rather than treating it as a generic software inventory item.

An extension with all-sites read access is an unreviewed integration with production data.