Skip to content

Security Policy Management

Security Policy Management is designed for platform administrators to configure AI gateway security policies per cluster. After policies are applied to the gateway plugin in a worker cluster, model requests and responses can be inspected and handled according to the configured action, such as blocking the request, logging only, returning a random replacement response, or masking matched content.

This capability depends on Higress integrated with the Knoway gateway. It requires LLM Studio v0.15.0 or later, with Higress enabled on the worker cluster. For how to enable Higress, see Upgrade Notes.

Prerequisites

  • The current user has platform administrator permissions and can access the Admin Console.
  • The target worker cluster has hydra-agent installed, and knoway.higress.enabled is turned on as described in the upgrade notes.
  • The policy list only shows clusters that already have the gateway security plugin configuration. If the list is empty, finish installing and upgrading Higress-related components and CRDs first.

Entry Point

  1. Sign in to the platform and open the Admin Console.
  2. In the left navigation, choose Security → Policy Management.

Policy management list

View the Policy List

The policy list shows gateway security configurations by cluster. The main fields are:

Field Description
Cluster Worker cluster name
Policy Types Currently enabled policy types; multiple types can be enabled together
Created At When the cluster policy configuration was created
Updated At When the configuration was last updated

You can search by cluster name in the search box. Click Edit on a row to open the policy editor for that cluster.

Note

There is no separate Create Policy entry. Policies are edited per cluster against an existing gateway plugin configuration. Clusters without Higress installation do not appear in the list.

Edit a Policy

  1. In the policy list, find the target cluster and click Edit.
  2. On the Edit Policy page, Cluster is read-only.
  3. Under Policy Configuration, enable the policies you need, fill in the parameters, and save.

Edit policy

Note

Enabled policies with configured parameters take effect. Disabled policies are not applied.

Policy Types and Fields

Policy Type Description Configurable Fields Actions
Regex Match Policy Match user input against regular expressions line by line to detect risk patterns such as jailbreak phrases and prompt injection Match Strategy: Trigger on any match / Trigger only when all match; Regex Patterns (one regex per line); when Random Replace is selected, also configure Random Replace Responses (one response per line) Block Request, Log, Random Replace, Mask Replace
Sensitive Word Policy Detect whether the input contains configured sensitive words for compliance and content safety Match Strategy: Trigger on any word hit / Trigger only when all words hit; Sensitive Words (one word per line); when Random Replace is selected, also configure Random Replace Responses Block Request, Log, Random Replace, Mask Replace
JSON Schema Validation Policy Validate the request body structure (field types, required fields, and so on) against a JSON Schema Enable toggle and Action. The UI does not provide online editing of the Schema body itself Block Request, Log
JWT Validation Policy Read a JWT from the specified request header and verify validity (signature, expiration, and so on) Header Name; Token Required: Required / Optional Block Request, Log
Model Allow / Block List Policy Allow only specified models via an allowlist, or deny specified models via a blocklist Mode: Allowlist (only models in the list are allowed) / Blocklist (models in the list are blocked); Select Models (multi-select enabled MaaS models in the cluster); Model Regex (one regex per line, useful for matching a model series) Block Request, Log

Action Descriptions

Action Description
Block Request Reject the call when the policy is matched
Log Forward the request normally and write a security audit log for later investigation
Random Replace Do not return the model output; randomly return one of the configured response texts
Mask Replace Return the content after masking the matched parts

Configuration Validation

Before saving, note the following validation rules:

  • When Regex Match Policy is enabled, provide at least one regex pattern.
  • When Sensitive Word Policy is enabled, provide at least one sensitive word.
  • When the action is Random Replace, provide at least one random replace response.
  • When Model Allow / Block List Policy is enabled, select at least one model or provide at least one model regex.

The edit page also provides Common examples that you can adapt for typical regex and sensitive-word patterns.

Notes

  • Policies take effect per cluster. Different worker clusters under the same platform are independent and must be edited separately.
  • Models available in the Model Allow / Block List come from enabled MaaS models in that cluster. Complete onboarding and enablement in MaaS Models first.
  • After you save policy changes, they apply to new gateway requests. Validate rules and actions in a test environment before enabling Block Request in production.
  • Triggered events can be queried and exported in Security Audit Logs.

Comments