Microsoft Wants Guardian Agents to Keep AI Threats in Check

Microsoft has warned that autonomous AI agents could become insider threats within organizations, even without malicious intent or external compromise. The company is calling for stronger security controls as businesses increasingly give AI systems access to sensitive data and internal tools.

In a new security proposal, Microsoft argues that AI agents can behave unpredictably, potentially exposing confidential information, accessing unauthorized resources, or taking actions that violate company policies.

Unlike traditional software, autonomous agents can make decisions and execute tasks independently, creating security risks that existing monitoring systems may struggle to detect.

Microsoft wants Guardian Agents to monitor AI activity

To address these risks, Microsoft proposes deploying Guardian Agents, specialized AI systems that monitor other agents in real time and intervene when they detect potentially harmful behavior.

The approach builds on Project Perception, Microsoft’s AI security initiative introduced in July 2026.

The framework combines sensors and signals to identify AI agents and track their activities across identity, cloud, network, endpoint, and data systems. It also uses contextual analysis to continuously assess agent behavior and identify potential security risks.

Guardian Agents then supervise these activities in real time, detect suspicious behavior, and intervene based on the severity of the threat. Microsoft believes these controls could help organizations prevent AI agents from exposing sensitive information or performing unauthorized operations.

Microsoft calls for greater transparency in AI models

Microsoft also wants AI developers to provide greater visibility into how autonomous agents make decisions.

The company argues that access to more information about AI reasoning and decision-making could help security systems distinguish legitimate tasks from potentially dangerous behavior.

This would allow organizations to determine whether an agent’s actions match its assigned responsibilities and internal security policies.

Microsoft is encouraging businesses to reassess how they manage AI agent permissions, monitor autonomous activity, and enforce security controls as agent adoption grows.

In other news, Microsoft and Carnegie Mellon launched the PRAISE AI training program, while Microsoft extended App-V Server support until October 2029.