AzureFeeds

At Microsoft Secure 2025, we introduced a new wave of innovations across Microsoft Defender aimed at redefining what AI can do for security operations. At the center of these announcements was the launch of 11 Security Copilot agents, each purpose-built to reduce manual workload and accelerate response through autonomous, adaptive automation. Integrated into existing Microsoft Security infrastructure, they continuously learn and adapt to your unique environment, while keeping your team in control for proactive, end-to-end protection.  

Among these is the Phishing Triage Agent in Microsoft Defender, now available in Public Preview. It tackles one of the most repetitive tasks in the SOC: handling reports of user-submitted phish. Instead of manually combing through endless submission, security teams can now rely on an agent that triages thousands of alerts each day, typically within 15 minutes of detection. Early adopters are already seeing accelerated threat response and significant time savings. 

Phishing: A top threat and a burden for SOC analysts 

Phishing continues to be one of the most pervasive entry points for threat actors, with over 90% of breaches starting from email-based deception. In just twelve months, Microsoft Defender for Office 365 detected more than 775 million malware-laced emails, underscoring the relentless, large-scale nature of the threat. While today’s security tools are highly effective at blocking most of these attempts, attackers aren’t standing still. They continuously adapt—tweaking content, spoofing identities, changing tactics, and exploiting new channels to slip past defenses. Increasingly, they’re also using generative AI to craft phishing messages that appear more legitimate and personalized, making detection even harder. As a result, a small but dangerous number of phishing emails still manage to slip through and reach users’ inboxes. 

When users report these suspicious messages, they land in SOC queues for further review, creating a significant operational burden for security teams. Most submissions are false alarms, yet analysts must still manually review each one to catch the rare threats buried in the noise. This delays response, drains focus, and raises the risk of a dangerous miss. 

Behind the agent: smarter phishing triage 

Built to operate autonomously 

The Phishing Triage Agent marks a meaningful step forward in autonomous security operations. Powered by large language models (LLMs), it performs sophisticated assessments—including semantic evaluation of email content, URL and file inspection, and intent detection—to determine whether a submission is a true phishing threat or a false alarm. Unlike traditional systems based on static rules or pre-coded logic, the agent dynamically interprets the context and artifacts of each email to reach an independent verdict. It is autonomous defense working behind the scenes, cutting through the noise and elevating what truly matters. 

Learning from feedback 

Equally transformative is the agent’s ability to learn. Rather than relying on fixed conclusions, the Phishing Triage Agent continuously evolves. Analysts can reclassify incidents and provide natural language feedback explaining why a particular verdict was correct or not. The agent incorporates this input, refining its reasoning and adapting to the organization’s specific needs, patterns, and nuances. With every interaction, it becomes more accurate and better attuned to its environment, creating a feedback loop that drives ongoing improvement. 

Transparent by design 

One of the most defining features of the Phishing Triage Agent is how clearly it communicates its decisions. For every verdict, the agent provides a natural language explanation that outlines why a message was or wasn’t classified as phishing. The rationale is clear and accessible, allowing analysts to quickly comprehend what led to the outcome. 

For those seeking deeper understanding, the agent also produces a visual map of its decision logic: a step-by-step breakdown of how it evaluated the submission. Each phase is presented as an expandable card within a structured diagram, detailing the signals analyzed, evidence collected, and logic applied. Teams can drill into any step to view the agent’s reasoning in context, making the entire process traceable and reviewable from start to finish. This level of transparency isn’t just helpful, it’s essential for building trust in autonomous security systems.  

How the agent works 

Quick setup and seamless integration 

Getting started is simple. The onboarding experience provides a clear overview of the agent’s capabilities and how it functions in your environment. It can be configured with a dedicated identity and role-based access controls that follow least privilege principles, ensuring it operates strictly within its assigned scope. 

Administrators retain full control. They can view, manage, and restrict the agent’s actions, keeping its behavior aligned with the organization’s security policies and standards.  

Autonomous operation in the background 

Once deployed, the agent operates in the background, automatically triggering whenever a user reports a suspicious email. As new submissions come in, it analyzes each one and assigns a classification. In most organizations, more than 90% of reported emails turn out to be false positives. The agent resolves these automatically, tagging them so analysts don’t have to sort through each one manually. This allows teams to focus on the handful of incidents that truly require their attention.  

When the agent is enabled, Microsoft Defender for Office 365’s built-in Automated Investigation and Response (AIR) feature consumes the agent’s output. AIR then builds on that analysis, detecting similar threats and surfacing remediation actions for SOC analysts to review and approve. 

Inside a typical incident review 

For each incident, the agent provides a natural language summary of its verdict. When it classifies a submission as malicious or benign, it clearly explains why—citing factors such as sender reputation, message content, attachment behavior, and more.  

The Activity tab displays a flow diagram that shows how the agent arrived at its decision, including all intermediate steps and outcomes. This behind-the-scenes process covers everything from text and URL analysis to sandbox evaluation of attachments. And it all happens autonomously, without human intervention or scripting. 

Feedback loop 

If an analyst disagrees with the agent’s verdict, they can simply reclassify the submission and leave feedback in natural language. No special syntax or training is required. The agent learns from this input and uses it to refine future decisions, continuously improving its accuracy and alignment. Over time, the agent becomes a true extension of the team. It not only reduces manual effort but also adapts to the organization’s unique environment and the evolving threat landscape it defends against. 

Visualized performance 

The agent’s performance is tracked in a dedicated dashboard that gives analysts real-time visibility into its impact. It displays the number of incidents handled, mean time to triage (MTTT), and a breakdown of false positives versus true positives over time. This always-on view helps security teams quantify efficiency gains, monitor accuracy, and build confidence in the agent’s ongoing performance. 

Responsible AI by default 

The Phishing Triage Agent, like all Microsoft Security Copilot agents, adheres to Microsoft’s Responsible AI principles. This includes built-in guardrails for fairness, transparency, security, privacy, and accountability. 

Administrators configure the agent’s identity and permissions based on least privilege access, maintaining strict control over what data it can access, how much capacity it consumes, and which actions it is authorized to take. Operating within a Zero Trust framework, the agent’s every action is evaluated against organizational policies before execution. This approach ensures that AI-powered capabilities enhance the SOC without compromising enterprise trust, compliance, or control. 

Supercharged SOC efficiency 

The Phishing Triage Agent is the first in a new generation of agents designed to bring autonomous intelligence to security operations. By eliminating repetitive, reactive tasks and continuously learning from feedback, it allows teams to focus on what matters most—investigating real threats and strengthening their overall security posture. This marks a leap forward into a more efficient, adaptive era for the SOC. 

Organizations that meet the prerequisites can now get started by joining the Phishing Triage Agent Public Preview, available through a trial directly in the Microsoft Defender portal. To learn more, visit the product page for details on how it works, and the Adoption Hub for broader guidance on Security Copilot agents. 

Looking to improve response times and support your team more effectively? Sign up to access “What generative AI can do for your SOC” today or read more about Microsoft AI-powered unified security operations.