Manage guardrails

How to add, configure, order, and save the guardrails on a Sentinel

Each Sentinel runs its own guardrail policy — the guardrails applied to inbound traffic, in the order you set. Admins manage that policy per gateway from the SUPERWISE® app, so different Sentinels can enforce different rules. Read-only users can view the policy but not change it.

Open a Sentinel and go to the Guardrails tab. The input pipeline lists the active guardrails in execution order, each showing its action and average added latency.

Add or remove a guardrail

Edit a guardrail

Select a guardrail to open its editor, where you set its behavior and parameters.

Action when triggered

Choose what happens when the guardrail matches:

  • Remediate — fix and forward. Each matched value is replaced with your Replacement text (default REDACTED, which appears in the prompt as {{REDACTED}}) and the request continues to the model.
  • Block — reject the request. Set the Error message returned to the caller.
  • Flag — record and forward unchanged. The match is logged as a finding, but the request and response are not modified. Use it to measure how often a guardrail would trigger before you turn on active enforcement — a safe way to try a new guardrail or a config change on real traffic. Flagged interactions appear in Live activity and the dashboards under findings.

Some guardrails only support one action. Language check, Image resolution, and Image sharpness, for example, can only block — there's nothing to redact.

Blocking and conversation history

⚠️

Sentinel checks each request on its own, over the whole prompt it receives — including the full conversation history. Agentic tools like Claude Code and the Codex CLI resend that entire history on every turn, so a message that triggered a guardrail once stays in the history and is sent again and again.

With a block action this can wedge a session: once a turn's content matches, every later turn replays it and is blocked too, even when the new message is fine. Remediate and flag don't have this problem — they let each turn through (remediate redacts the match; flag only records it). So for content that can recur in history, prefer remediate or flag over block, and if a session does get stuck on a blocked guardrail, start a new conversation so the triggering message is no longer in the history.

Parameters

Some guardrails expose parameters:

  • PII redaction lets you choose the detected categories — pick exactly which identifier types to scan for (see PII redaction).
  • PHI detection lets you choose the detected categories and a confidence threshold for the detection model (see PHI detection).
  • Language check lets you set the allowed (or blocked) languages and a confidence threshold (see Language check).
  • Image resolution lets you set a maximum or minimum and the width × height limit in pixels (see Image resolution).
  • Image sharpness lets you set a minimum sharpness score below which an image is blocked as too blurry (see Image sharpness).

Test sandbox

Before you commit a guardrail's settings, test them in the editor's Test sandbox. Provide a sample request — text for a text guardrail, or an image for Image resolution and Image sharpness — click Run test, and Sentinel evaluates it against the guardrail's staged configuration and shows the outcome:

  • Passes — the request wouldn't be touched.
  • Would remediate — the matched content, with your replacement text applied, so you can see exactly what the model would receive.
  • Would block — the request would be rejected, with the error message the caller would get.

The sandbox runs the check in the control plane and doesn't affect live traffic or telemetry, so you can iterate on parameters and the action until the outcome is what you want, then apply and save.

Generate test cases

Instead of writing sample inputs by hand, you can have Sentinel generate example test cases from the guardrail's current configuration — so you can see how a change to the config would affect what the guardrail catches. In the editor, click Generate examples from this rule: Sentinel suggests a set of relevant samples (text for a text guardrail, images for an image guardrail), tailored to the settings you've staged.

  • Try an example to run it straight through the Test sandbox and see the outcome.
  • Regenerate a single example, or Regenerate all, to get a fresh set — useful after you change a parameter.

This is the quickest way to check that a config change does what you expect before you apply and save.

Order the pipeline

Guardrails run in order on each request. Drag a guardrail to reposition it in the input pipeline — for example, to redact secrets before running other checks.

Apply and save

Edits are staged in two steps:

  1. In a guardrail's editor, click Apply changes to stage them.
  2. On the Sentinel page, click Save changes to publish the policy — or Discard to revert.
ℹ️

Saved changes take effect on the Sentinel's next heartbeat — gateways pull their policy from the control plane about once a minute. A guardrail introduced in a newer release also needs the gateway to be running a current image; if a guardrail you added doesn't take effect, see Keep your Sentinel up to date.

🔒

Editing a Sentinel's guardrails is an admin action. Read-only users can view the policy but can't change it.

Get Sentinel: Create a free SUPERWISE account.

See also


Did this page help you?