Guardrails

Taimoe Enterprise AI Gateway provides Guardrails to sanitize LLM requests before reaching upstream model providers and inspect responses before returning them to client applications.

With Guardrails, security and operations teams can centrally define safety policies and rules, enforce content sanitization, and monitor real-time security events across all gateway traffic.

---

Policies & Rules Management

Guardrail policies are scoped per organization and team. Each policy bundles one or more content safety rules.

Creating a New Policy

1. Navigate to GOVERNANCE → Guardrails in the Aegis Console.
2. In the Policies & Rules tab, click "New Policy" in the top right.
3. Configure the policy details:
   • Policy Name: Name of the policy (e.g., Production-Safety-Policy).
   • Description: Brief explanation of the policy's scope (e.g., Content safety rules enforced on production LLM traffic).
4. Add Guardrail Rules:
   • Cloud Provider: Guardrail safety engine (currently supports GCP Model Armor).
   • Credential: Select the corresponding cloud credential (gcp provider type).
   • Resource Reference (resource_ref): Enter the resource reference path for the cloud safety template (e.g., GCP Model Armor Template Resource Path).
   • Apply To: Direction to apply the rule:
     • input: Inspect request prompts only.
     • output: Inspect LLM responses only.
     • both: Inspect both request prompts and model responses.
   • Mode: Rule enforcement mode:
     • enforce: Immediately blocks violating requests or sanitizes output, returning a 400 or 403 HTTP status code.
     • observe: Logs safety violations to Audit Logs without blocking original requests (ideal for testing new policies in staging).
5. Click "Save" to persist the policy.

Editing & Deleting Policies

• Edit Policy: Click the Edit icon in the policy table row to update policy names, descriptions, or modify rule order and parameters.
• Delete Policy: Click the Delete icon in the row action menu to remove a single policy.
• Bulk Delete: Select multiple checkboxes in the table and click "Delete Selected (N)" in the top toolbar to remove policies in bulk.

---

Security Events Monitoring

The Security Events tab provides a real-time audit view of requests blocked or flagged by guardrail rules.

1. Switch to the Security Events tab.
2. The event log automatically streams events with WARNING and ERROR severity.

Event Log Fields

Field	Description
Time	Exact timestamp when the security event occurred.
Severity	Severity level of the event (WARNING or ERROR).
Action	Action or API route triggered (e.g., /v1/chat/completions).
Model	Model alias targeted by the request.
HTTP	HTTP response status code returned to the client.
Message	Detailed error message or rule violation details returned by the safety engine (e.g., prompt injection detection or PII flag).

Search & Filtering

Use the search bar at the top of the table to filter security events by Action, Model, or Message in real time to locate potential security threats.