Source: Microsoft Foundry — Guardrails and Controls Overview

Foundry's guardrails (formerly content filters): risks, four intervention points including agent tool calls, blocking levels, defaults and override rules.

Contents

Summary

Explains how a guardrail (a named set of controls) detects risks at chosen intervention points and either annotates or blocks.

Key claims

  • Risks: hate, sexual, self-harm, violence, user prompt attacks, indirect attacks, spotlighting (preview), protected material (text/code), groundedness (preview), PII (preview) and task adherence (preview).
  • Intervention points: user input, output, and, for agents in preview, tool call and tool response.
  • Blocking levels: Off, Low (blocks only high severity), Medium, High (blocks low and up). Turning output filtering off requires approval for modified guardrails.
  • Models get Microsoft.DefaultV2 by default. An agent’s guardrail fully overrides its model’s; without one it inherits the model deployment’s.
  • Groundedness detection in Azure OpenAI retires April 30, 2027.
  • Guardrails apply only to agents built in Foundry Agent Service, not to other agents registered in the control plane. Hosted agents can add network egress controls (preview).

Feeds

Azure AI Foundry

Suggest changes →