Detect and investigate threats to AI agents using Microsoft Defender (Preview)

Deployed AI agents operate autonomously, invoking tools, accessing data, and taking actions on connected systems in response to natural‑language input. This makes continuous detection and investigation critical. Microsoft Defender detects suspicious and malicious agent behavior, provides alerts in near‑real‑time, and enables security teams to investigate incidents and trace the full root cause and blast radius.

Microsoft Defender detects and enables security teams to investigate threats to AI agents managed through Microsoft Agent 365, including the extended detection capabilities available for supported agent platforms. To block unsafe agent actions during runtime, see Protect AI agents in real time using Microsoft Defender.

Prerequisites

• Enable security for AI agents, including the Microsoft 365 app connector to collect Agent 365 observability data for AI agent actions. See Enable security for AI agents using Microsoft Defender.
• Ensure that your AI agent emits observability data to Microsoft 365:
  • Agents built with Microsoft Copilot Studio, Microsoft Foundry, and declarative agents built with the Microsoft 365 Copilot Agent Builder send observability data to Microsoft 365 by default.
  • For AI agents built on other platforms, enable observability using the Microsoft Agent 365 SDK, as described in the Agent 365 development lifecycle documentation.
• To detect threats to local AI agents that run on endpoints, set up AI agent runtime protection in Microsoft Defender for Endpoint. Microsoft Defender for Endpoint must run in active mode. Local agents are onboarded separately from cloud agents.
• (Optional) To include the prompt snippets that triggered a detection as evidence in alerts, enable prompt evidence collection. This setting is enabled by default.

Detect AI agent threats in near-real-time

Microsoft Defender continuously monitors AI agent activity and detects suspicious and malicious behavior for all Agent 365‑managed agents. Defender analyzes agent telemetry, tool usage, and execution patterns to identify threats such as jailbreak attempts, XPIA attempts, malicious content propagation, secrets and credentials leak, evasion techniques, suspicious user access, and more.

Microsoft Defender surfaces detections as near‑real‑time alerts in the Defender portal and enables security teams to investigate them using familiar security operations workflows, including alert triage, incident correlation, and Advanced Hunting.

For more information, see Incidents and alerts in the Microsoft Defender portal.

Near-real-time detections rely on Agent 365 observability data, which also provides valuable context for investigating incidents and threat hunting. Microsoft Defender analyzes this data to identify suspicious agent behavior and generate alerts.

To enrich alert investigation with the prompt snippets that triggered a detection, enable prompt evidence collection. For more information, see Control prompt evidence in alerts.

Investigate AI agent threats and hunt for risks using Advanced Hunting

Microsoft Defender correlates AI agent alerts into incidents and surfaces the related context so security teams can quickly assess impact and prioritize response. Advanced Hunting then lets analysts query Agent 365 observability data by using Kusto Query Language (KQL) to investigate incidents and hunt for risks throughout their environment.

Investigate incidents and alerts

Microsoft Defender correlates AI agent alerts from near‑real‑time detections into incidents. Real‑time protection audit and block events are recorded as behaviors in the BehaviorInfo table, which you can correlate with alerts during investigation.

Security analysts can use the incident graph and investigation experience to understand the full context of a potential attack, including relationships between involved entities and the blast radius of AI agent threats. For more information, see Incidents and alerts in the Microsoft Defender portal.

Correlate alerts and Agent 365 observability data and hunt for risks using Advanced Hunting

Advanced Hunting in Microsoft Defender enables security teams to query Agent 365 observability data alongside other security data by using Kusto Query Language (KQL). This supports proactive threat hunting, incident investigation, and root‑cause analysis for agents, applications, identities, and devices.

For example, use Advanced Hunting to:

• Trace specific agent tool invocations and correlate them with related alerts or block events

• Investigate the root cause and scope of a detected AI agent threat

• Identify anomalous execution patterns or risky agent behavior throughout environments

• Build custom detection rules based on agent activity signals

Advanced Hunting tables for AI agent investigation

The following Advanced Hunting tables provide visibility into AI agent configuration, alerts, and activity. You can query these tables individually or correlate them to investigate incidents and hunt for agent-related risks.

Next steps

• Protect AI agents in real time using Microsoft Defender

• Discover AI agents and assess security posture using Microsoft Defender

• Protect AI assets from emerging threats and vulnerabilities using Microsoft Defender