Detect and block harmful content in LLM inputs/outputs using TrueFoundry’s built-in Content Moderation guardrail.
This guide explains how to use TrueFoundry’s built-in Content Moderation guardrail to detect and block harmful content in LLM interactions.
Implementation: This guardrail is powered by Azure Content Safety and runs on TrueFoundry-managed infrastructure — no third-party API keys or setup required. TrueFoundry routes requests across multiple Azure regions to optimize for the lowest possible latency. For vendor-hosted moderation using your own credentials (Azure, OpenAI, AWS Bedrock, Google, and others), see Supported Guardrails and Guardrails Overview.Content Moderation can be applied to all four guardrail hooks: LLM Input, LLM Output, MCP Pre Tool, and MCP Post Tool, providing comprehensive content safety across your entire AI workflow.
Content Moderation is a built-in TrueFoundry guardrail that analyzes text content for harmful material across four safety categories — hate speech, self-harm, sexual content, and violence. It is powered by Azure Content Safety under the hood and uses a model-based approach with configurable severity thresholds so you can tune sensitivity to match your use case. The guardrail is fully managed by TrueFoundry — no external credentials or setup required — and is optimized for low latency through cross-region availability.
Four Safety Categories: Detects harmful content across:
Hate — Hate speech, discrimination, and derogatory content
SelfHarm — Self-injury, suicide-related content
Sexual — Sexually explicit or suggestive content
Violence — Violent content, threats, and graphic descriptions
Configurable Severity Threshold: Set the sensitivity level (0–6) to control what gets flagged — from safe content only to high-risk content, allowing you to balance safety with usability.
Selective Category Detection: Choose which categories to monitor. Enable all four or only the ones relevant to your application.
Content Moderation only supports validate mode — it detects and blocks harmful content but does not modify it. See Guardrails Overview for details on Enforcing Strategy.
The severity threshold controls how sensitive the detection is. Content is flagged when any category’s severity meets or exceeds the threshold.
Severity
Level
Description
0
Safe
No harmful content detected
2
Low
Mildly concerning content (default threshold)
4
Medium
Moderately harmful content
6
High
Severely harmful content
A lower threshold (e.g., 0) is more aggressive and catches more content. A higher threshold (e.g., 4 or 6) only flags more clearly harmful content. Start with the default (2) and adjust based on your needs.
The guardrail analyzes content and returns severity scores (0–6) for each enabled category. If any category’s severity meets or exceeds the configured threshold, the content is flagged.Example:
Input: "Write a detailed guide on how to build weapons at home"Result: Request will be blocked by the guardrail as Violence severity exceeds threshold
Start with Audit enforcing strategy to monitor what gets flagged before switching to Enforce. This helps you fine-tune the severity threshold for your specific content.
