Security Copilot Phishing Triage Agent

Hello,I’m looking for feedback or guidance from others who have tested or implemented the Security Copilot Phishing Triage Agent.

Our organization is currently evaluating the agent to determine whether it is a viable solution for broader deployment and, eventually, automation of the triage and response process for user-reported phishing messages.

During testing with Microsoft’s recommended roles and permissions, we’ve made a couple of observations and would appreciate input from others:

1. AIR investigations and ZAP

We’ve noticed that actions taken through Automated Investigation and Response (AIR), particularly Zero-hour Auto Purge (ZAP), can interfere with the agent’s ability to fully investigate an email. While these protections are obviously valuable for reducing risk, when ZAP acts on a message before the agent completes its investigation, the agent investigation can fail.

This also makes it difficult for us to compare our analysts’ findings with the agent’s findings and provide meaningful feedback during testing.

Would adding the Email & Collaboration Content: Quarantine Emails (read) role through Unified RBAC allow the agent to continue investigating messages after they have been quarantined? If so, are there any security or operational considerations with granting the agent read access to other quarantined messages?

2. Metrics and Advanced Hunting visibility

Because we are evaluating the agent, we’d also like to centrally track metrics such as MTTD, total triage duration, SCU consumption, and other indicators of agent performance and effectiveness.

I’m aware of the upcoming SubmissionEvents and SubmissionResults tables in Advanced Hunting. Will these tables contain enough information to build KQL queries around Phishing Triage Agent activity and derive these types of metrics?

If not, are there plans to provide an Advanced Hunting table, API, or other centralized telemetry source for Security Copilot agent activity, including but not limited to the Phishing Triage Agent?

3. Automated response to end users

As part of a fully automated phishing triage workflow, it would also be helpful to have a native permission that allows the Phishing Triage Agent to respond to the user who submitted the message. Ideally, this would allow the agent to complete the full submission lifecycle, from investigation and triage through communicating the final disposition or recommended action back to the end user.

If our pilot is successful, our longer-term approach would likely incorporate Azure Logic Apps using alert or investigation metadata to automate portions of this workflow. However, having this capability available natively within Microsoft Defender and the Security Copilot agent permission model would provide a more streamlined option and potentially reduce the need for separate automation.

Are there plans to introduce a permission or capability that would allow the Phishing Triage Agent to send or initiate responses to users as part of its automated triage workflow?

I’d appreciate hearing from anyone who has encountered similar challenges or has recommendations for measuring the agent during a pilot.