Skip to main content
Some sites dress up as a well-known AI service to collect sign-ins or to push a fake desktop app. They are impersonating sites, also called lookalike sites or fake AI sites. Their front pages often say exactly what the real service says, so the evidence that would normally confirm an AI service would confirm the fake one too. Blindsight therefore asks a different question first: not “is this AI”, but “is this pretending to be an AI service it is not”. Every host your fleet reaches is checked against the names of the major AI services before anything else is decided about it. This page covers what happens when one matches, what is blocked, and what you can change.

What Blindsight does with a lookalike

A host that matches is listed on Endpoint → Shadow AI as Impersonates service, for example Impersonates ChatGPT: the real service is chatgpt.com. From then on:
  • It is never treated as an AI service. It is not counted in the Shadow AI figures, on the Overview’s What changed, in report findings, or in In use, Observed, Inspected, Approved and Waiting on the AI services page.
  • It ranks High risk, with the real service named.
  • It is never approved or acted on by the Newly discovered AI setting, and never suggested for the shared catalogue.
  • It raises the Impersonating site reached alert, once per site.
  • Approve is unavailable on its row. The row tells you which domain to approve if people need the real service.
The real service is never flagged. A brand’s own domains, every host in the built-in catalogue and every built-in rule are never matched.

How a lookalike is recognised

Expand the row to see which check matched. Only the host name is read. Blindsight does not publish its list of service names or its matching rules.

Which lookalikes are blocked

Blocking is a setting of your workspace, and it only touches the clearest cases:
  • a site spelled with lookalike characters
  • a download or sign-in word beside the service’s name
  • a near miss of the name, or the name on a different ending, when other evidence supports it, such as how recently the domain was registered
Every other lookalike is named and alerted, never blocked. A site you decided about yourself is never touched by this setting. Block impersonating sites is on for workspaces created from 16 November 2026. It is off for a workspace that existed before, until an administrator turns it on. There are two places to do it:
  • the notice that says how many sites your people reached that pretend to be a well-known AI service, with Turn on blocking. It shows at the top of every page of the console, to the people who can edit the Default policy, and stays until you act on it or close it.
  • Endpoint → Policies → Default policy, the Block impersonating sites switch
A workspace that has blocking on without ever having chosen it is told, while the block holds hosts, that blocking is on and where to turn it off. Every change is written to the audit trail.

Block or warn

Under When a person opens one, choose what a device does with a site the rule covers:
  • Block refuses the site.
  • Warn shows a page saying the site is not the service it imitates, with a link to the real service. The person may continue anyway, and the site then opens on that one device for a short time. Each continue is recorded and counted on the row (Opened N times past the warning in the last 30 days).
Warn needs desktop 2.1.8 or later. An older desktop app blocks. Moving from Block to Warn weakens a protection, so it is recorded at high severity in the audit trail and raises the policy-weakened alert. A site you blocked yourself is always refused.

What people see

  • A person who is blocked sees a block page that names the service the site copies and where the real one is (desktop 2.1.7 and later). Workspaces can add a short message and a contact to it; see Block page message and contact.
  • In a workspace that names people, each person who reached a blocked lookalike gets one plain email, once per site, naming the site, the service it imitates and the real address. It tells them to report it if they typed anything there. A weaker lookalike is never emailed, and a workspace that shows totals only sends nothing.

What to do with one

  1. Open the row on Endpoint → Shadow AI and read which check matched.
  2. If blocking is off, choose Block on the row, or turn on Block impersonating sites for all of them.
  3. Check the devices and people listed against it. Anyone who reached it may have entered work credentials or run a download.
  4. If people need the real service, approve its real domain.

Not a lookalike

If a site is listed by mistake, choose Not a lookalike on its row. It is offered on rows nobody has decided about yet. Devices then stop treating the site as a copy of that service. The site is not blocked and not read: this is a record of your decision, not an approval, and the row reads Not a lookalike rather than Approved. Each decision is written to the audit trail.

The alert

Impersonating site reached is a high-severity alert sent once per site. Choose it at Endpoint → Configurations → Alerts like any other alert. In a workspace that shows totals only, the alert names the site and nothing about who reached it or when.