Skip to main content
Discovery only is one switch for a whole workspace. It makes your devices watch and nothing else: they find which AI services people use, nothing is blocked, no prompt is changed, and no prompt text is sent. It is also called watch only or passive mode. Use it for a first look before you protect anything, for a proof of concept, or while a works council or a legal review is still open. It is off until an administrator turns it on, and turning it off puts back what it changed.

What it changes

Turning it on moves these settings together, in one audited change. Only the settings that are not already at the value below move. The certificate, the local proxy and the network adapter stay installed. Only the host names of AI services are recorded. A file’s name is never kept, and the attachment’s type and size are. See Protections for what each setting does on its own.

Personal data found on the laptop

On a device with desktop app 2.1.12 or later, Discovery only can also look for personal data without acting on it. The laptop reads the prompt on the machine only to find the data types your workspace chose, and reports the category and a count. It changes nothing, blocks nothing and sends no prompt text. A stored record never holds the text.
  • It needs your plan to cover personal data protection.
  • It applies only to devices that are new enough and report that they can do it. An older device records host names only. The Discovery only card says how many of your devices record personal data, for example 3 of 5 devices record personal data (category and count only). The rest cannot until their desktop app is updated.
  • A data type you set to Keep is not looked for.
  • In Overview, a prompt that a device checked and kept on the laptop has its own outcome in the traffic flow, Kept on the laptop, which is not a gap in inspection.

Turn it on

You need the Endpoint licence, Shadow AI in your plan, and the endpoint.protections.configure permission. If turning it on would also move Block impersonating sites, you need runtime_security.firewall.configure too. There are two places:
  • Endpoint → Policies → Default policy → Overview, the Discovery only card, button Make this workspace discovery only. Save or discard any other changes on the page first, so the settings you see are the ones that change.
  • Setup, on the Devices step, What should the agent do?, then Watch only (passive). Setup uses the same switch. When devices are already checking in, Setup asks before it switches them, and says how many.
The dialog lists, one per line, what stops: prompts no longer masked, prompts no longer checked for injection, Unapproved AI no longer protected or blocked, and lookalike sites no longer blocked. Devices pick it up at their next check-in.

On, and in force

Discovery only is in force only when nothing else would still make a device read, block or send. If something would, the card says Discovery only is on, but not in force and lists each cause with a link to where it is set:
  • Shadow AI is off, or not in your plan.
  • Prompt injection detection or personal data protection is on again.
  • Unapproved AI is not Record only, or Block impersonating sites is on.
  • An AI service is set to inspect, or a named policy reads or blocks, or a service is blocked.
  • An app rule refuses an application, or refuses applications that cannot be named.
  • Someone changed one of the moved settings by hand after you turned it on.
Setup says Watching only only when Discovery only is on and in force.

Turn it off

Choose Turn off discovery only. Each setting it moved goes back to what it was. These are left as they are, and the dialog says which:
  • A setting someone changed by hand while Discovery only was on.
  • A setting your plan does not cover.
  • Prompt injection detection, which stays off while cloud inspection is off.
  • A setting whose restoring would leave no protection on at all, because at least one protection has to stay on.

What is recorded

  • Turning it on or off writes one settings change to the audit trail, with each setting’s value before and after.
  • Moving prompt injection detection to off, and back, has its own record.
  • Each setting turned down raises one policy weakened alert, so one switch can raise several.