Ai Agent Runtime Protection Overview
In brief
The page now documents inaccurate blocked-alert titles in Audit mode and a Claude Code limitation where tool responses may continue after a Block-mode alert.
What Defender admins need to know
Check the configured protection mode and investigate detections rather than relying on alert titles. Post-tool-response blocks cannot undo actions already completed.
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.
For local PowerShell and organization-wide Microsoft Intune configuration options, see Set up AI agent runtime protection with Microsoft Defender for Endpoint.
Known issues (Preview)
Important
During preview, prompt injection alerts in the Microsoft Defender portal might incorrectly report Suspicious AI prompt injection blocked when AI agent runtime protection is configured in Audit mode. This is an alert-reporting issue. In Audit mode, runtime protection allows the action to continue and records the detection; the alert title doesn't mean that runtime protection blocked the action.
Verify the runtime protection mode on the affected device as described in Set up AI agent runtime protection. Don't rely on the alert title alone to determine whether an action was blocked. Continue investigating the detection to assess whether the activity is malicious.
Important
During preview, Claude Code might continue processing a tool response even when AI agent runtime protection is configured in Block mode and Defender reports a post-tool-response block. This is a blocking limitation, not just an alert-reporting issue. On affected devices, the block alert doesn't mean that runtime protection prevented the agent from processing the detected content.
Don't rely on the alert title alone to determine whether a tool response was blocked. Continue investigating the detection to assess whether the activity is malicious. Post-tool-response inspection occurs after the tool runs and can't undo actions the tool has already completed.
Investigation
When runtime protection detects prompt injection, Defender raises a Suspicious AI prompt injection alert and correlates related activity into incidents for investigation.
@@ -9,8 +9,8 @@ feedback_product_url: https://techcommunity.microsoft.com/t5/security-compliance uhfHeaderId: MSDocsHeader-MicrosoftDefender manager: bagol description: Learn how Microsoft Defender for Endpoint provides runtime protection for local AI agents by detecting and blocking prompt injection attacks.-author: lwainstein-ms.author: lwainstein+author: j0shbregman+ms.author: joshbregman ms.service: defender-endpoint ms.topic: overview ms.custom: msecd-doc-authoring-1030@@ -102,6 +102,20 @@ Microsoft recommends starting in audit mode to observe detections and validate a
For local PowerShell and organization-wide Microsoft Intune configuration options, see [Set up AI agent runtime protection with Microsoft Defender for Endpoint](configure-ai-agent-runtime-protection).
+## Known issues (Preview)
+
+Important
+
+During preview, prompt injection alerts in the Microsoft Defender portal might incorrectly report **Suspicious AI prompt injection blocked** when AI agent runtime protection is configured in **Audit** mode. This is an alert-reporting issue. In Audit mode, runtime protection allows the action to continue and records the detection; the alert title doesn't mean that runtime protection blocked the action.
+
+Verify the runtime protection mode on the affected device as described in [Set up AI agent runtime protection](configure-ai-agent-runtime-protection#enable-runtime-protection). Don't rely on the alert title alone to determine whether an action was blocked. Continue investigating the detection to assess whether the activity is malicious.
+
+Important
+
+During preview, Claude Code might continue processing a tool response even when AI agent runtime protection is configured in **Block** mode and Defender reports a post-tool-response block. This is a blocking limitation, not just an alert-reporting issue. On affected devices, the block alert doesn't mean that runtime protection prevented the agent from processing the detected content.
+
+Don't rely on the alert title alone to determine whether a tool response was blocked. Continue investigating the detection to assess whether the activity is malicious. Post-tool-response inspection occurs after the tool runs and can't undo actions the tool has already completed.
+
## Investigation
When runtime protection detects prompt injection, Defender raises a **Suspicious AI prompt injection** alert and correlates related activity into incidents for investigation.
