MP.AIGOV.1046 Azure AI model deployments should use a content filter that blocks prompt attacks
Prompt Shields block prompts written to make the model ignore its rules. Without a content filter that blocks them, a user or attacker can jailbreak the model into producing harmful output or acting against your instructions.
Remediation action
In the Foundry portal, go to your project, select Build > Guardrails > Create guardrail. Add a
User prompt attack control with the Annotate and block action (and a Document attack control if
your app passes documents or tool results to the model). Assign the guardrail to each model deployment. In
Foundry (classic) the same setting is Guardrails + controls > Content filters > Prompt Shields for
jailbreak. A guardrail is stored as the deployment's raiPolicyName.
Related links
- Configure guardrails and controls in Microsoft Foundry
- Configure content filters
- Prompt Shields
- MCSB v2: Artificial Intelligence Security (AI-2)
- How this check works
%TestResult%