A workspace where nothing has been seen opens on Catalogue. An empty
Observed tab reads “No AI traffic observed yet. Agents report each service the
first time it is used.” with an Open the Catalogue link. On a workspace that
shows only counts it reads “No AI traffic released yet. A service is listed here
once a closed week releases its counts.”
A workspace with
traffic opens on Observed. The count beside each tab is the number of rows in
it. Search by host or vendor.
On a workspace that shows only counts, Observed lists closed weeks. See
Who is named in the console.
On a platform that does not serve the Catalogue yet, the page keeps its single
list without tabs. Nothing is shown as 0 when the platform could not say.
The four decisions
Each row offers the same decisions, on both tabs. They needendpoint.ai_services.manage. Without it the row says Needs the manage grant.
Some services are also Managed apps: a service you onboarded as an app. A
managed app is the strongest decision, so Approve and Block are off for it.
Remove the app itself to stop managing it.
The console uses the words Inspect and Protect the other way round from
older pages. Inspect is the light posture: nobody is blocked. Protect is
the strong one, where a request that trips a rule can be stopped. The next pages
use the console’s words.
Approving keeps inspection on
Approving a service does not switch inspection off. Approve only removes the service from Shadow AI and from the unapproved-AI block. Blocking is its own decision. On an approved service, sensitive data is pseudonymized or redacted, not blocked. A block stays possible for reasons that are not about personal data: prompt injection, harmful content, or an explicit block. An approved row reads:
A row with no decision reads Not decided, inspected, Not decided, counted
only or Not decided plus what the default does with it.
What is inspected by default
Built-in services and the most-used AI services (popular ones) are inspected by default, without anyone approving or deciding. Hosts that cannot be intercepted, such as cloud, identity and operating system update hosts, read counted only and the row says these settings have no effect on them. Approving a service removes it from Shadow AI and keeps it inspected. Services approved before this release are moved to inspected by the release’s data migration, except those where an admin chose Stop inspecting: those keep the service settings Off for pseudonymization and prompt-injection detection. Approving or inspecting such a service later does not turn the detectors back on. The row says so: “Approved, but Pseudonymization is Off for this service … Approving does not turn it back on: change it here.”Microsoft 365 Copilot
Copilot is decided as one family. A decision oncopilot.cloud.microsoft also
covers m365copilotapp.svc.cloud.microsoft. The row says Also covers and
names the other host. Clear removes the whole family. A rule you set by hand on
one of the family hosts stops the decision with a message that names it.
The catalogue grows
Blindsight keeps a shared catalogue of AI services. It is read live, so a service added to it is recognised in your workspace without an update on your side. More AI services are recognised over time, and a service you added by hand keeps its own record alongside.Shadow AI shows unapproved services only
The Shadow AI page lists recognised AI that is not approved. An approved or managed service lives on AI services. The figures on the Shadow AI page are Unapproved AI, Waiting on you, High risk and Blocked. When every recognised service your agents saw is approved, the page reads No unapproved AI and points you to the AI services page. Waiting on you counts services where nothing is decided and nothing is inspecting them. A page that cannot tell reads Not available, not 0.Per-service settings: Customize for this service
Under a service, Customize for this service opens settings that apply to that service only. While it is closed, a line says what the service follows. The settings are:- Personal data: Inherit, Off, Redact or Pseudonymize.
- Prompt injection: Inherit, Off or On.
- Data types: Change a data type to set Protect, Keep or Block for one kind of value in prompts to this service. See Protections.
Which setting wins
The first setting that is not Inherit wins, in this order (the same order applies to data types):- The policy’s setting for this service.
- The service’s own setting.
- The policy’s default.
- The workspace setting.
endpoint.ai_services.manage. A person
without the permission sees the rows read-only: “Changing this needs the manage
grant.” Each Catalogue and Observed row opens an inspection panel for the host
that governs it, with these controls and the line “Policies can change this for
the people they reach.”
The service’s setting applies to the host whose rule governs the service. A
setting on example.com covers the subdomains its rule covers. A custom service
with an exact host covers that host only. A host no rule governs is refused.
Settings in a policy
A policy has an Inspection section with Customize. It holds:- Default for every service in this policy: Personal data (Inherit, Off, Redact or Pseudonymize), Prompt injection (Inherit, Off or On) and Data types. A default applies to inspected services, not to blocked or tunnelled ones.
- Single services: a list of services, each with its own settings for that policy.
endpoint.protections.configure, which an admin holds
by default. Without it the section is read-only, and a save that changes it is
refused with a message naming the permission. The rest of the policy is not saved
in that case either.
Works council workspaces
In a workspace with the works council module on, switching prompt-injection detection On is refused until the works council notice flow exists, whether from a policy or for one service. Admins see:This workspace has a works council. Prompt-injection detection cannot be switched on until the notice flow exists; ask support. An On already stored stays, and the policy can be saved with it.In the policy editor the On choice for prompt-injection detection is disabled, with the hint “Needs the notice flow in a works-council workspace”, except where that exact On is already stored. For a service, the change is refused with
422 works_council_notice_required. Off, Inherit, pseudonymization
On and saving an unchanged value are never refused.

