Guardrails Configuration

Input/output sensitivity thresholds and persistence

Guardrails Configuration

This guide covers the Guardrails Dashboard in the Admin Console, including sensitivity configuration, event monitoring, and false positive reporting.

Overview

The platform enforces input and output guardrails at Layer 1 (Experience & Interaction) of the architecture. Every message sent to an agent and every response generated by an agent passes through the guardrails proxy, which checks for safety violations before the content reaches the LLM or the user.

The Guardrails Dashboard is located in the Admin Console at /guardrails.

Guardrail Types

The platform detects six categories of safety violations:

Each detected event has a severity level: low, medium, high, or critical.

Blocked vs. Warning

When a guardrail is triggered:

• Blocked — the message is prevented from reaching the LLM (input) or the user (output). The agent or user receives a safety notice.
• Warning — the message is allowed through but flagged for review. The event is logged for monitoring.

Whether a trigger results in a block or warning depends on the severity level and the configured sensitivity threshold.

Sensitivity Configuration

Accessing Configuration

1. Open the Admin Console and navigate to Guardrails.
2. Expand the Configuration section (collapsed by default).

Input Guardrail Sensitivity

Controls the PII detection threshold for incoming user messages.

• Slider range: 0% (lenient) to 100% (strict), in 5% increments.
• Lower values — fewer messages are blocked; only high-confidence PII detections trigger.
• Higher values — more aggressive blocking; even partial matches are flagged.
• Default: 75%.

Output Guardrail Sensitivity

Controls the toxicity and hallucination detection threshold for agent responses.

• Slider range: 0% (lenient) to 100% (strict), in 5% increments.
• Lower values — only severe toxicity or clear hallucinations are blocked.
• Higher values — stricter filtering; borderline content is blocked.
• Default: 60%.

Saving Configuration

1. Adjust the sliders to the desired levels.
2. Click Save Configuration.
3. A confirmation message appears on success.
4. The configuration is persisted to the tenant's guardrail config via PUT /api/admin/guardrails/config.

Configuration changes take effect immediately for all agents in the tenant.

What the Guardrails Proxy Does

The guardrails proxy sits between the user/agent and the LLM. The flow is:

User message
  --> Input guardrails check (PII, prompt injection, jailbreak, off-topic)
    --> LLM inference (via AIFS)
      --> Output guardrails check (hallucination, toxicity)
        --> Response delivered to user

For n8n agents, the guardrails proxy intercepts LLM calls as a sidecar:

• n8n's OpenAI credential points to http://localhost:8081/v2 (the proxy).
• The proxy applies input/output checks.
• Then forwards to the AIFS endpoint for inference.

For platform-managed agents, guardrails are applied inline in the chat handler before and after the LLM call.

Dashboard Metrics

The top of the Guardrails page shows summary cards:

Charts

Trigger Rate (7-Day Trend)

A line chart showing daily trigger counts by guardrail type over the last 7 days. Each guardrail type has its own colored line, allowing you to spot trends (e.g., increasing prompt injection attempts).

Input vs. Output Guardrails

A bar chart comparing the number of input-side vs. output-side events.

Events by Trigger Type

A horizontal bar chart ranking guardrail types by event count, from most to least triggered.

Event Log

Below the charts, a filterable Guardrail Event Log table lists individual events.

Filters

• Type — Input, Output, or All.
• Trigger — PII Detection, Prompt Injection, Jailbreak, Hallucination, Toxicity, Off-Topic, or All.
• Severity — Low, Medium, High, Critical, or All.
• Blocked Only — toggle to show only blocked events.
• Clear filters — resets all filters to defaults.

The table shows the matching event count (e.g., "Showing 42 of 156 events") and supports pagination (25 events per page).

Table Columns

Event Detail Panel

Click any row to open the detail panel (right side on desktop), which shows:

• Event ID.
• Agent name and ID.
• Type and trigger badges.
• Severity level with color indicator.
• Trigger Explanation — a detailed description of the specific guardrail type and what it detects.
• Full content preview.
• Status (Blocked/Warning) and false positive badge if marked.
• Timestamp (absolute and relative).
• Action button to mark/unmark as false positive.

False Positive Reporting

False positive reporting helps improve guardrail accuracy over time.

Marking a False Positive

There are two ways to mark an event as a false positive:

1. From the table — click the flag icon in the FP column. The icon turns blue when marked.
2. From the detail panel — click "Mark as False Positive". The button changes to "Unmark False Positive" once marked.

What Happens When You Report

• The event is flagged with falsePositive: true via POST /api/guardrails/{id}/false-positive.
• The False Positives metric on the dashboard is updated.
• Over time, false positive data is used to tune guardrail sensitivity and reduce incorrect blocks.
• You can unmark a false positive at any time by clicking the flag icon again.

Related Guides

• Admin Guide — general Admin Console navigation
• Security and Compliance — broader platform security features