Microsoft Defender for Endpoint
Endpoint protection

Configure Ai Agent Runtime Protection

In brief

The page now states that sensitive data and secrets may appear unredacted in prompt snippets shown in alerts. It also links to current Prompt Evidence Collection details and notes that Intune template help text is being finalized during preview.

What Defender admins need to know

Administrators should account for potential sensitive data exposure when reviewing or sharing alerts.

Summaries are generated from the documentation change itself.

Documentation change

The comparison below shows only the changed extract. Use the full-page view for complete context.

| Ai Agent Network Inspection (AiAgentNetworkInspection) | Inspect supported agent-to-LLM network traffic. | Choose audit or block mode independently of agent-native event inspection. The default is disabled. | | Prompt Evidence Collection (PromptEvidenceCollection) | Collect prompt evidence when AI agent runtime protection is configured. | Enable or disable collection. |

Important

Sensitive data and secrets aren't redacted from prompt snippets shown in alerts. During preview, the help text in the Intune template is being finalized. For current details about Prompt Evidence Collection, see Prompt evidence collection.

For each inspection method, Audit detects and reports prompt injection while allowing the action to continue, and Block blocks detected threats at supported inspection points. Leave a method disabled if you don't need it. You can enable either inspection method or both.

Microsoft recommends that you initially assign the policy in audit mode to a limited device group. Review detections, and then change the inspection settings to block mode when you're ready to enforce protection. Choose whether to enable prompt evidence collection separately. Complete the policy wizard, assign the policy to the device groups you want to protect, and then create the policy.