Skip to main content

What users see and findings

Updated on 10 November 2026.

When an Endpoint Agent policy applies to a request, the person at the device sees a response in their AI app or coding tool, and QuilrAI records the decision in Observe > Findings & Interactions.

Key terms​

TermMeaning
AI appAn AI chat or assistant app, such as ChatGPT, in a browser or on the desktop.
Coding toolAn AI coding agent on the device, such as Claude Code or Cursor.
ResponseWhat a rule tells the agent to do. See Responses.
InteractionOne recorded request, such as a prompt or an upload.
FindingA recorded request that needs review. See What makes a finding.

What people see in AI apps​

The agent checks each prompt and upload before it reaches the AI service. With no decision from QuilrAI within 5 seconds, the request is sent. See When a device can't get a decision.

  • Block stops the request.
  • Require a reason holds the request until the person gives a reason.
  • Ask for a reason sends the request and shows a reason popup.
  • Record only sends the request with no message.

Block​

The AI service never receives the request. Apps whose error format QuilrAI knows show their own block message, and no QuilrAI popup appears. Other apps receive an HTTP 406 error with the text "Blocked by DLP policy" (DLP is data loss prevention), and QuilrAI shows a Prompt Blocked popup asking the person to remove confidential information or contact the IT helpdesk. It closes when dismissed or after 30 minutes.

Require a reason​

  1. The Potentially Sensitive Data Detected. popup lists the data types found and previews the content.
  2. Under Reason for bypassing:, the person picks a listed reason, such as False Positive, or writes 20 to 140 characters under Other (write your own)….
  3. The person selects Continue with Original within the 15-second Time left: countdown, which is set on the device, not in the Console.

The request is then sent, with the reason on the finding. It is blocked if the countdown ends, or if the popup is closed or cannot be shown.

If the countdown ends, the popup stays open and says the time to justify has elapsed. A reason given there within 10 minutes is recorded as a second finding. It does not unblock the request, but the same content sent again within 2 minutes goes through without a new popup.

Popups appear one at a time, with a counter such as 1 of 3. A Require a reason popup that waits more than 15 seconds behind others is dropped, and its request is blocked.

Ask for a reason​

The request is sent at once, and the same popup appears without a countdown. It closes by itself after 30 minutes. The finding is recorded when the person answers or closes it, so it can appear later than the request.

Account notices​

When an account, plan or model is not approved under Accounts, the agent refuses the request and shows Workspace/Org ID not approved, Plan not approved or Model not approved. See Files, Skills, Packages, Apps and Accounts.

What people see in coding tools​

The agent checks the coding tools in Differences between coding tools. A rule set to All coding tools covers every coding tool the agent checks, including tools supported later. With no decision within 7.5 seconds, the prompt or action continues.

  • Block stops the prompt or action. A tool that cannot stop a step shows a warning instead.
  • Warn and continue (Ask for a reason in AI apps) shows a warning, and the prompt or action continues.
  • Record only shows nothing.

Require a reason is not available for coding tools. Use Warn and continue instead.

The message in the tool​

Each message starts with QuilrAI Guardian Agent, for example "QuilrAI Guardian Agent blocked your prompt". For a prompt or a file read, it names the data types found. For a file type or sensitivity label check, a skill or a package install, it names the item and the reason. It never shows the detected values.

Differences between coding tools​

Coding toolPrompts checkedDesktop popup on Block
Claude Code, Qoder, QoderWorkYesFor prompts
Cursor, Codex CLIYesNo
Copilot CLI, Copilot in the IDEOnly files referenced in the prompt. Coding-tool rules don't see typed text. To check it, add a rule for the GitHub Copilot app under AI apps.No
AntigravityNo. Tool actions are checked.No
ZCodeYesFor prompts and tool actions

The desktop popup is the Prompt Blocked popup with the tool's message. Warnings never show it. In Antigravity, Warn and continue asks the person to approve the action.

Responses kept from earlier settings​

  • Notify and continue is no longer offered. Rules saved with it still work: AI apps show a Security Notice popup and let the person continue but record a block, and coding tools block. Choose another response.
  • Rules converted from the per-app Endpoint settings can use a redaction response. In AI apps it blocks the request, because the agent does not rewrite requests.

While enforcement is paused​

Rules are still checked, but Sensitive data and Files responses become Record only, so people see no policy blocks, popups or warnings. Findings are marked Paused. Some controls, such as Apps and account checks in AI apps, keep applying. See What changes while enforcement is paused.

How decisions appear in Findings & Interactions​

To see only Endpoint Agent records, set the Sensor filter to Endpoint Agent. See Findings & Interactions.

Go toQuilrAI ConsoleObserveFindings & Interactions

In AI apps, the Endpoint Agent reports every request it inspects. A report that fails to send three times, or is still waiting to send when the agent stops or restarts, is dropped, so that request does not appear. In coding tools, an allowed check with no sensitive data is not recorded.

What makes a finding​

A request becomes a finding when any of these is true:

  • The response was Block, Require a reason, Ask for a reason or Warn and continue.
  • A Record only rule that looks at content matched: it found a data type, or checks a file, command, tool action or MCP condition.
  • Sensitive data was found, but no published rule matched it.
  • The Policy Engine could not check a request with sensitive data.

Every other request is an interaction only, including one decided by an allow rule: an exception with endpoint.action = allow in a policy's Code view (see Edit a policy as code). Record only rules are not allow rules.

Where the deciding policy is shown​

The Cards layout names the policy and version in the finding sentence. The Table layout's Policy column adds the rule's own risk. The drawer, which opens when you select a finding, shows the policy and any reason the person gave. Open policy opens the policy, or History at the deciding version if it is no longer live.

The policy shows its current name. If it was renamed, hover over it to see its name in the deciding version. People who cannot read Endpoint Agent policies see the name without a link.

What each policy label means​

LabelMeaning
Policy name Version 12A rule in this policy decided, on version 12 of your published policies.
+2 more after the nameOther policies gave the same response or set the risk. Hover to see their names.
Allowed by policy nameAn allow rule decided. The request is an interaction only.
Paused · would have blockedEnforcement was paused, and the named rule would have blocked. Other responses read the same way, for example Paused · would have required a reason.
Paused · would-have action not recordedEnforcement was paused. The policies that matched are listed after Matched:, without the action they would have taken.
PausedEnforcement was paused, and no rule is credited. Paused is also added after No rule matched this data and Rules not checked.
No rule matched this dataSensitive data was found, but no published rule matched it. Create a policy starts one.
Rules not checked, Some rules not checkedSee Why rules were not checked.
Legacy Endpoint settingsThe per-app Endpoint settings decided, while the Policy Engine was off or for a coding-tool file type or sensitivity label check.
Policy Engine V1: rule nameThe earlier Policy Engine (Detection Policy) decided.
Decision label: nameThe agent's name for the decision, without recording what decided it.

Why rules were not checked​

Hover over the label for the reason:

  • Rules not checked (no rule was checked): the sensitive data scan didn't finish, your policies couldn't be loaded or confirmed, an error occurred, or the request was too large.
  • Some rules not checked (only some rules were skipped): data types your rules need weren't available, or the request was too large for every rule.

Rule risk and Decision risk​

Risk comes from each rule's Risk level on findings field and never changes the response. See How risk is set.

  • Rule risk is the risk on the rule the finding names, for example Rule risk: High or No risk set on this rule. Older findings can show Rule risk: Not recorded.
  • Decision risk is the highest risk any matching rule set, from any policy. The drawer shows it only when it differs from the rule risk, or No risk set when no rule set one.

Severity​

Severity is a finding's triage priority: Critical, High, Medium or Low. It comes from the first source that applies:

  1. A severity set by hand.
  2. Policy risk, while it is on and a matching rule set a risk: the Decision risk, even if a severity rule would rate it higher.
  3. The first matching severity rule.
  4. The default severity.

Policy risk is off by default. To turn it on, you need permission to triage findings across your whole organization.

  1. In Findings & Interactions, select Triage center, then Severity & due dates.
  2. On the Policy risk card, turn on Policy risk sets severity.
  3. In Let policy risk set severity?, select Turn on.

Open findings are classified again, which can change their severity and due dates. See Triage center.

Until a new finding gets its severity, usually within minutes, the Cards and Table layouts show its Decision risk in place of a severity. Filters and due dates do not use that value. The drawer, and findings with no Decision risk, show Pending.

In the drawer, hover over the severity to see its source, such as Set by email, Rule: rule name or Policy risk: High · policy name.

Export decisions​

Use Export on Findings & Interactions, or Export Center. Allow-rule requests appear in Interactions exports only. Policy columns include:

  • Policy: the policy the finding names.
  • Policy · Outcome: rule, paused_rule, allowed_by_rule, no_policy (no rule matched) or not_evaluated (rules not checked).
  • Policy · Rule Title: the policy's title in the deciding version.
  • Policy · Risk Level (Decision risk) and Policy · Rule Risk Level (Rule risk). If Policy · Risk Level Recorded is true, a blank risk means no rule set one.

Each column also appears a second time with Activity · in front of its name, for example Activity · Policy · Outcome (Policy becomes Activity · Policy · Name). Values that were not recorded are blank.

Activity on each policy​

In Govern > Policy Engine, choose the Endpoint Agent workspace, then Desktop > Controls > All policies. The Last 7 days column summarizes each policy, and its Overview tab has details.

  • Lost to another policy counts matches where one of its rules was stricter but a higher-priority policy won. It stays 0 while all policies keep the default priority.
  • Top person and Model most blocked always show Not measured yet.

Days are UTC calendar days. A coverage line under the tiles, such as Based on 2,448 of 4,965 checked requests (49%) since 31 Oct, shows how many requests the counts include. Only devices that report every decision are counted, from the day recording started. Other requests were still checked.

Notes and limits​

  • Test with new content each time. The same content sent to an AI app again within about 2 minutes can get no new popup or finding. See Behavior that can change a result without a message.
  • The Browser Extension can check the same prompt. In a browser that both products monitor, each applies its own policies, so one prompt can show two popups and record two findings.
  • Reasons stay on the finding. They are not sent to Action requests, which handles Browser Extension requests only.
  • Apps with Application Monitoring off stop recording new activity, with the exceptions in Application Monitoring is off for the app. Existing findings stay.