Skip to main content

How controls are evaluated

Updated on 10 November 2026.

This page explains how the Browser Extension picks the one control that applies to a user's action in the browser, and what changes the result. You manage controls under Govern > Policy Engine > Endpoint Agent > Browser; see Browser controls.

Go toQuilrAI ConsoleGovernPolicy EngineEndpoint AgentBrowser

Key terms​

TermMeaning
ControlOne browser rule: the activity to watch for, conditions, a mode and actions.
System controlA control that QuilrAI supplies, tagged System. Controls that admins create are Custom.
Use caseThe activity a control watches for, such as A user is pasting sensitive data. It sets the mandatory conditions and the risk behavior.
Risk behaviorThe kind of risky activity a use case belongs to, such as Users Pasting Data. Each has a fixed rank. See the Use case reference.
CriticalityVery High to Very Low. It breaks ties between controls of the same rank.
ModeMonitor records a match without interrupting the user. Enforce applies the control's Just in Time action.
Just in Time actionThe only action type that runs in the browser as the user acts, such as a block, a Justify popup or a Prompt notice. Other action types run on the QuilrAI backend. See What users see.
Browser UtilityA desktop program installed with the extension on Windows and macOS, also called the Browser Agent. It supplies the content of copied text and uploaded files.

How the extension decides, step by step​

  1. Download the controls. At startup and every 2 minutes, the extension downloads every enabled System and Custom control.
  2. Detect the action. It checks only controls whose use case covers that kind of action, so a paste control is not checked when the user submits a prompt. A post into a website chatbot can also match Users Leaking Sensitive Data controls. See Which control wins when several match.
  3. Check the conditions. A control matches when all of its conditions are true. QuilrAI checks content for sensitive data using your Detection Models.
  4. Remove excused controls. It drops controls excused for this app or person, for example by a turned-off detection or an approved action request, and controls whose popup has reached its prompt frequency limit.
  5. Pick one winner. It ranks the rest by risk behavior, then by criticality, and applies only the first.
  6. Apply the winner.
    • Monitor: the action goes ahead and a finding is recorded.
    • Enforce: the Just in Time action applies, unless the user already gave a reason or acknowledged a notice for the same risk behavior in that tab and site. See What changes the outcome for an app or a person.
    • Enforce without a Just in Time action: the user is not interrupted. Chromium-based browsers and Firefox record a finding, and Safari records nothing.

The finding in Findings & Interactions names the winning control.

How conditions are checked​

Inside a group, rows combine from top to bottom, and AND does not take precedence over OR. See How conditions combine.

Join top-level condition cards with AND

Every browser ignores a control whose top-level condition cards are joined with OR, although the console saves it and shows it as enabled. Put the alternatives inside one group card. See OR between top-level condition cards.

Which control wins when several match​

The extension ranks matching controls in two steps:

  1. Risk behavior rank. It is fixed, and you cannot change it. Rank 1 is the highest.
  2. Criticality. Within the same rank, the higher criticality wins: Very High, High, Moderate, Low, Very Low.

If two controls tie on both, the winner is not defined. Give overlapping controls different criticality values.

When the user...Risk behaviors checked, by rank
Opens a page1 Blocked AI Applications; 2 Blocked SaaS Applications; 3 Unapproved AI Applications Used, Visiting SaaS app, Newly Introduced AI Apps and Newly Introduced AI Users
Signs in or signs up1 SignIn and SignUp to Applications; 2 Unapproved AI Applications Used; 3 Credential Sharing; 4 Password Reuse; 5 Compromised Password; 6 Weak Password
Posts into a website chatbot1 Users Posting Data to a Chatbot; 2 Users Leaking Sensitive Data, when the post is checked for sensitive data
Submits a promptUsers Posting Data or Users Leaking Sensitive Data, never both (see below)
Copies, pastes or uploadsOne risk behavior each (Users Copying Data, Users Pasting Data, Users Uploading Data), so criticality alone decides

To find the risk behavior of a use case, see the Use case reference.

For a prompt:

  • If a Users Posting Data control applies to the user on that app, the extension applies it, even in Monitor, without checking for sensitive data. Users Leaking Sensitive Data controls do not run there.
  • Otherwise, the extension checks the prompt for sensitive data, and criticality decides between the matching Users Leaking Sensitive Data controls.

Do not use both risk behaviors for the same people and apps.

A Monitor control can silence an Enforce control​

warning

The extension picks the winner before it looks at the mode. If the winner is in Monitor, the action goes ahead and is only recorded, even when a lower-ranked Enforce control would have blocked it. To make sure an Enforce control applies:

  • Give it a higher criticality than every Monitor control of the same rank that matches the same action.
  • Narrow or turn off any higher-ranked Monitor control that matches the same action. Criticality cannot overcome rank.

System controls compete too​

System controls are ranked like Custom controls. Ten are on by default, all in Monitor. Most match only on Generative AI apps.

ControlRisk behaviorCriticality
Preventing The Leakage Of Sensitive Data Via An AI AppUsers Leaking Sensitive DataModerate
Tracking Usage Of An Unapproved AI AppUnapproved AI Applications UsedModerate
Discovering New AI AppNewly Introduced AI AppsLow
Discovering a new AI userNewly Introduced AI UsersLow
User is copying sensitive dataUsers Copying DataLow
User is pasting sensitive dataUsers Pasting DataLow
copy-paste data from one app to anotherUsers Pasting DataLow
User is trying to upload sensitive dataUsers Uploading DataLow
Logging into SaaS applications using shared credentials of another non admin userCredential SharingLow
Logging into SaaS applications using credentials of another admin userCredential SharingVery Low

For example, a Custom Enforce control on the same use case as Preventing The Leakage Of Sensitive Data Via An AI App needs High or Very High to win. At Moderate the winner is not defined. Below that, the System control wins and users see no popup.

Against a higher-ranked System control, criticality cannot help: narrow or turn off the System control. Tracking Usage Of An Unapproved AI App matches sign-ins and sign-ups (not visits) on unapproved AI apps, so it outranks every password and shared-credential control there. Filter the controls list by Type to see System controls.

Apps you mark Blocked in Inventory are blocked by built-in controls that are hidden from the controls list. They run in every browser at Low criticality and record each visit. Do not add your own control on AI application blocked by organisation or SaaS application blocked by organisation. It cannot block, because these use cases have no Just in Time action. Above Low, it outranks the built-in block and unblocks the users it matches.

What changes the outcome for an app or a person​

Most of these settings remove controls before ranking (step 4), so a lower-ranked control can win instead. A reason or acknowledgement given on a popup applies after ranking, to the winner only (step 6).

SettingWhere you set itEffect
A detection turned off for an appInventory > app > Controls & configuration > Detection behaviorsRemoves every control of that risk behavior on the app, for every user.
An approved action requestGovern > Action RequestsRemoves every control for that action (for example Upload), including its content checks, for one person on one app.
A reason given on a Justify popup, or a Prompt notice acknowledgedIn the browserSkips the popup once, the next time the same risk behavior wins in that tab and site. For password and shared-credential controls, the pass lasts until the tab closes, for the same account and password. It does not apply to Users Copying Data, Users Uploading Data or Users Posting Data controls, or to a reason given in the sensitive data window.
Site on the allowed domains listSettings > Browser Extension > Whitelist Domain & PatternThe extension does not run on the site, for any user.

How prompt frequency affects the result​

Prompt frequency, a setting on some Prompt and Justify actions, limits how often the popup appears. For its options, see Conditions and actions. The extension:

  • Applies it only to controls on page visits and sign-ins. For other actions, the popup appears every time the control wins.
  • Keeps the count per app. Controls with the same risk behavior share one count on that app.
  • Counts separately in each browser and each installation.
  • Removes a hidden control before ranking (step 4), so a lower-ranked control can win instead.

When nothing matches​

If no control matches, or step 4 removed every match, the action goes ahead. The user sees nothing, and no finding names a control. If you expected a control to apply, check that:

  • It is enabled, was saved more than 2 minutes ago, and joins its top-level condition cards with AND.
  • No Monitor control (System controls included) or Users Posting Data control takes precedence over it.
  • Step 4 did not remove it (a turned-off detection, an approved action request or prompt frequency), and the site is not on Whitelist Domain & Pattern.
  • An earlier reason or acknowledgement in that tab did not skip the popup.
  • The extension is on for the user and has linked the browser to that user. See Extension settings.
  • For copied text and file content, the Browser Utility is running on the device. See Content checks need the Browser Utility.

Notes and limits​

  • Timeouts and offline. If a content check is slow or QuilrAI cannot be reached, the action goes ahead and the content is treated as not sensitive. If the extension cannot download controls, it keeps its last controls. See When a check takes too long and When a browser cannot reach QuilrAI.
  • Sign-up forms. While the user types a password on a sign-up form, the extension checks only Unapproved AI Applications Used, Password Reuse, Compromised Password and Weak Password controls, and an Enforce winner shows only a notice. Its own action applies when the user submits the form.
  • Change timing. A save takes effect in QuilrAI at once. There are no drafts or versions. Browsers pick up a control change within about 2 minutes, and an app setting or action request decision usually within about 5 minutes. See How long changes take.
  • The save check is not a ranking. When you save, the console refuses a control whose conditions are equivalent to another control's in the same use case, including System and turned-off controls. This check ignores rank, mode and criticality. See Resolve a conflict.