Configure rules for the browser
Decide per guardrail what a person sees in the Browser Guard, set the outage behavior, and find the results.
The Browser Guard does not have its own rule set. It uses the guardrails of your project, the same ones that apply to SDKs, the gateway proxy and the MCP gateway. For each guardrail you decide in addition what the person sees in the browser when it applies.
Where you set it
- Open Guardrails in the sidebar (section Agents & Policies).
- On the My Guardrails tab, open the guardrail.
- Under Enforcement Action, the action of the guardrail is set for all paths.
- Under AI Browser Guard Action, you decide what happens in the browser.
Follow the action or choose your own
By default, Inherit from action is switched on. The browser then follows the Enforcement Action of the guardrail:
| Enforcement Action | In the browser |
|---|---|
| Block | The prompt is not sent. |
| Flag | Only when an external AI model checks the request: the person sees a warning and decides whether to send. When the gateway checks with its own models, the guardrail only records a hit; nothing happens in the browser then. |
| Justify or approve | The person writes a justification and can then send. Via SDK and API the same guardrail waits for an admin’s approval instead. |
| Anonymize | The prompt goes out masked, where the service allows this; there is no additional action in the browser. |
| Logs only | When the gateway checks with its own models: no action in the browser, the prompt goes out. When an external AI model checks the request, the guardrail blocks it on a hit. |
Switch Inherit from action off to choose one of these browser actions instead:
| Browser action | What the person sees |
|---|---|
| Block | The prompt is not sent; the person sees that it was blocked. |
| Warn | A warning; the person can choose to send anyway. |
| Justify | The person writes a reason before the prompt goes out. |
| Redirect | The person is pointed to an approved alternative. Enter its Alternative name and Alternative URL. |
What goes out is always exactly the content that was checked.
A guardrail that is decided by a specialist model (Prompt Guard 2) shows the hint that action, severity and browser action are not configurable there.
Justification for personal data
When the gateway masked personal data in a prompt from the browser, the person can send it with a justification. Which categories can then go in clear text, and which settings under Settings → Security keep company names and places masked, is described in the Trace Explorer and in Security settings.
Long prompts
The detection by models (personal data in running text, manipulation attempts, rules written in words) today reads only the beginning of a prompt, roughly 2,100 to 2,400 characters of German text. Keep this in mind for rules that rely on these detections; see the overview.
Behavior during an outage
If the gateway cannot check a prompt, the outage default of the project applies. You set it under Settings → Security → Browser Guard → Behavior during an outage:
- Let through (default): requests go out unchecked. They are marked visibly and reported afterwards, as soon as Palveron is reachable again.
- Block: nothing goes out. Your staff cannot send prompts in AI chats during this time.
A browser policy set by your IT (browser_outage_mode) takes precedence over this setting. Only admins can change it.
Rules for AI services
Whether an AI service counts as approved, you decide in Shadow AI by classifying it (see Evaluate Shadow AI). For an AI service that is not approved, the Browser Guard shows a warning; the person can continue, and continuing is reported. If the rules the Browser Guard loaded last are older than allowed (default 240 minutes, set through device management with policy_max_age_minutes), it shows a notice that the rules are not up to date.
Where you see the results
- Trace Explorer: each check from the browser, with what the browser reported about sending, and a justification or a send after a warning in the Human decision card (see Trace Explorer).
- Shadow AI, tab Governance: the table Recent Browser Activity with the column Transmission.
- Shadow AI, tab Reports: the Shadow AI report (CSV), if your plan includes the export.