Shadow AI Discovery
Overview
Shadow AI Discovery in HYPR AgentPass™ shows which AI software-as-a-service (SaaS) tools your workforce reaches through the browser. It is the first tab of Oversight → Discovered.
Enrolled workstations watch outbound connections against a HYPR-curated catalog of AI tools, and report what they see. From that, you can approve the tools your organization sanctions, block the ones it does not, and ignore the ones that were never shadow AI.
The second tab, Discovering Unmanaged Agents, observes unmanaged inference clients, at the gateway or by process on the workstation. Shadow AI Discovery observes SaaS domain access on the endpoint. Both are shadow AI, but they are different signals, and a tool can appear in both.
Before You Start
These must be true before any data appears:
- Agent Observability is enabled. Devices receive the list of domains to watch over the same signed channel they use for Agent Observability, so nothing reaches them while it is off. See Enabling Agent Observability.
- Devices are enrolled, and running an endpoint client that supports AI SaaS observation.
- Observation is switched on for your tenant, with the Observation toggle in the coverage card at the top of the page. It stays off until you turn it on.
Until then, the page shows zeros and says why. That is deliberate, so a screen of zeros is never mistaken for "nobody is using AI."
Who Can See and Act
Opening the tab needs permission to view agents. Every change on it needs permission to edit settings: approving, blocking or ignoring a tool, adding a tool and switching observation on or off. By default only administrators can edit settings, so operators can read the page but not change it.
The People tab, and the per-person detail behind it, also need the AI SaaS user detail permission. See Attribution and Privacy.
Reading the Page
Every figure on the page covers the last 30 days. The window is fixed.
Coverage Comes First
Every count on this page is only as complete as the fleet reporting it, so coverage is at the top, not in a footnote:
- Devices reporting / enrolled shows how much of your estate the page is based on. A device counts as reporting when it has reported in the last 7 days. If a third of your machines are not reporting, every number below understates reality by roughly a third.
- Attribution shows whether usage is attributed to named individuals, or aggregated. See Attribution and privacy.
- Warnings appear here when coverage is poor enough to mislead, for example when fewer than half of enrolled devices have reported in the last 7 days.
The Headline Numbers
Four figures summarize the estate.
| Figure | What it means |
|---|---|
| Tools in use / watched | How many catalog tools were actually reached, against how many your devices watch for. Ten out of 2,700 is a quiet estate. Ten out of eighteen is a catalog that is barely looking |
| Total accesses | Every observed access, sanctioned or not |
| Unsanctioned accesses | Accesses to tools you have not approved |
| Blocked | Accesses your endpoints refused |
Approved tools still count toward total usage. Both numbers matter, because the total gives the unsanctioned figure its context.
The Tabs
The page has four tabs:
- Tools: the 20 most-used tools, most-used first, with your decision on each. The count on the tab includes every tool in use.
- Watched catalog: everything your devices are matched against, whether or not anyone used it. It answers "what do you actually look for?" See The Watched Catalog.
- Workstations: usage by machine. Expand a row to see what that machine reached.
- People: usage by person, ordered by unsanctioned activity, not total. Somebody who uses a tool you approved heavily is not shadow AI, and should not top this list. Expand a row to see every tool the person reached, each tagged Approved, Blocked or Unsanctioned.
The Watched Catalog
The Watched catalog tab lists every tool your enrolled devices watch for, 25 to a page. Tools you added yourself are listed first on every page.
Use these controls to narrow the list:
- Search name or domain: find a tool by its name or one of its domains
- All categories: show one category only, such as Assistant, Code or Writing
- Any risk: show only High risk, Medium or Low tools
- Show ignored: include the tools you ignored
Each row shows a tool and your decision on it:
| Column | What it shows |
|---|---|
| Tool | The tool's name, with its description or category beneath. Yours marks a tool you added, and Ignored a tool you ignored |
| Domains matched | The domains your devices match for the tool. A dash means another tool holds the domain, so nothing matches on it here |
| Your decision | Approved, Blocked or a dash for no decision |
| Actions | Approve or Unapprove, Ignore and, for a tool that is not blocked, Block…. An ignored tool shows Stop ignoring instead |
Deciding About a Tool
A tool you have made no decision about is Unsanctioned. You can give it one of three decisions. They answer different questions, and only one applies at a time.
You decide from the tool's row in the Tools tab or the Watched catalog. Every decision except a block takes effect without a confirmation, and devices apply it on their next poll.
Approve
"We sanction this."
Approved tools are excluded from the unsanctioned figure, and cannot be blocked. Approval and a block would contradict each other, so approving clears any block.
Approved tools are still observed, but reported at a coarser resolution. You keep the totals, and your endpoints send far less traffic, since nobody needs a minute-by-minute breakdown of an assistant they approved.
Block
"We forbid this."
Blocking stops the connection on enrolled devices, and counts the refusals. Before it commits, AgentPass shows you what the block affects:
- Which domains it covers
- How many devices it affects
- Recent access volume
- Whether any other tool claims the same domain
The last point is why blocking is never a simple toggle. Dozens of entries in an AI directory can claim a single domain. github.com, for example, is listed as the home of 82 different AI tools.
To block a tool:
- Click Block… on the tool's row.
- Review the confirmation dialog. It shows the Devices affected and the Accesses (30 days), lists each domain under Domains this covers and warns about anything the block also touches.
- Click Block it. Devices apply the block on their next poll.
Each domain in the dialog carries a tag that says what the block does to it:
| Tag | Meaning |
|---|---|
| will be blocked | The block covers this domain |
| also claimed by N | Other tools claim the same domain, and blocking it affects them too |
| protected — will not be blocked | The domain is shared infrastructure, and stays unblocked |
| not blockable — still under review | The domain is still under review, and stays unblocked |
When a tool cannot be blocked, the button reads Cannot block, and the dialog says why: the tool is approved, it has no registered domain or none of its domains can be blocked.
Shared infrastructure, such as cloud, CDN and hosting domains, is protected, and cannot be blocked by any route. Blocking one would break unrelated services across your whole organization, not only the AI tool that happens to sit on it. The dialog tags the domain as protected, and the block covers the tool's other domains. When none of a tool's domains can be blocked, the dialog refuses.
Ignore
"This was never shadow AI."
Use it for a corporate tool that happens to sit in an AI directory, or a false positive on a shared domain.
Ignoring is stronger than hiding a row. The tool is dropped from what your devices watch for, so they stop matching and reporting it entirely, with no traffic and no stored rows. History already collected is hidden, not deleted. If you stop ignoring the tool, it comes back.
Approving a noisy row makes it go away, but misstates your figures. Approved usage counts as sanctioned AI, so approving something that is not really an AI tool overstates how much sanctioned AI your organization uses. Ignore is the accurate control.
Ignored tools disappear from the Tools tab, and from the Watched catalog by default. Select Show ignored to see and reverse them.
Reverse a Decision
Reverse a decision from the tool's row:
- Remove an approval. Click Unapprove in the Tools tab or the Watched catalog. The tool returns to Unsanctioned.
- Lift a block. Click Unblock in the Tools tab. That tab lists the 20 most-used tools of the last 30 days. For a blocked tool it does not list, click Approve in the Watched catalog, because approving clears the block. Click Unapprove afterwards to return the tool to no decision.
- Stop ignoring a tool. In the Watched catalog, select Show ignored, then click Stop ignoring on the tool. Devices watch for it again on their next poll. Ignoring cleared any approval or block, so decide about the tool again.
Adding a Tool of Your Own
The HYPR catalog cannot know about your internal tools, and no directory is complete. To add a tool:
- In the Watched catalog, click Add a tool. The Add a tool to watch dialog opens.
- Enter a Name, which is what administrators will read.
- Enter the Web address. AgentPass takes the hostname from it, and fetches and stores the tool's favicon as its logo. The logo is never loaded from the third-party site when the page is viewed.
- Under What is it?, write a sentence for whoever looks at this next.
- Click Add tool.
Your tool's domain reaches your enrolled devices on their next poll, and the tool appears in the Watched catalog marked Yours. It matches the domain and every subdomain of it.
Expect two deliberate refusals:
- The domain is already in the catalog. A domain identifies exactly one tool. AgentPass names the existing tool, so you can approve or block that instead.
- The domain is shared infrastructure. This is the same protection as for blocking, for the same reason.
If no logo is found, the tool is still added, and the message says so. The address may not publish a favicon.
Attribution and Privacy
Per-person usage is governed separately from the aggregate views:
- Attribution mode. In aggregate-only mode, no usage is ever attributed to an individual, and none is stored. The People tab then says it is empty by configuration, not because nobody is using AI tools. To turn on aggregate-only mode for your organization, contact HYPR support.
- A separate permission. Seeing named individuals requires the AI SaaS user detail permission, granted separately from the aggregate views. By default only administrators hold it. Users without it see the tools and the workstations, but not the people.
Detailed access records are kept for a short window. The aggregated figures behind the dashboard and its trend are kept far longer.
Where the Catalog Comes From
AgentPass ships with a curated catalog of several thousand AI tools, so discovery works on day one with no configuration and no internet access.
HYPR curates that catalog centrally. A tool's domain is delivered to every customer's endpoints, so entries are reviewed before they become active. The decision is yours, per tenant: approved, blocked or ignored. Your decisions are never visible to another customer, and theirs are never visible to you.
On-premises and air-gapped installations can load an updated catalog from a signed bundle. Your own gateway always serves the logos, which are never fetched from third-party sites when a page is viewed. Loading them from external URLs would tell whoever hosts those images which AI tools your workforce uses.
Troubleshooting
Start with what you see.
| What you see | What it means |
|---|---|
| "AI SaaS discovery is switched off for this tenant" | Observation is off, and the figures shown are historical. Use the Observation toggle |
| Zeros, and a low devices-reporting count | Few devices are reporting. Check enrollment, and that the endpoint client supports AI SaaS observation |
| A tool you expect is missing from the Watched catalog | Only active catalog entries are watched. Entries HYPR is still reviewing are not delivered. You can add the tool yourself |
| A blocked tool still shows accesses | Blocks apply on each device's next poll, not instantly. Accesses before then are already recorded |
| The People tab is empty | Either attribution is aggregate-only, or you lack the AI SaaS user detail permission. The page says which |