Microsoft adds Guardian Agents to monitor the agents inside a company
Microsoft is expanding Project Perception to monitor the AI agents companies deploy themselves. Guardian Agents accompany controls that restrict access and adjust permissions as tasks change. The company argues that unpredictable model behavior requires supervision separate from the model provider. People retain responsibility for policies and consequential decisions as automated tools gain access to sensitive business systems.
Artificial Intelligence··Evening
Guardian Agents watch the tools a company deploys
Microsoft has added Guardian Agents to Project Perception, its enterprise security product, to supervise AI agents operating inside a company. The October 19 announcement extends a system introduced in June. Its earlier agents had separate roles: one group simulated attackers, another detected attacks and assessed vulnerabilities, and a third worked on repairs. The new focus includes software that an organization has chosen to deploy and granted access to its own systems. Microsoft describes these tools as potential insiders, including when their behavior is unpredictable rather than deliberately malicious. Guardian Agents are intended to monitor their activity within boundaries set by people.[1], [2]
Permissions change with the task
Other announced controls restrict the data and tools available to an agent and adjust its permissions during a task. Risky decisions can require human approval. Microsoft has also added agent discovery to Defender, its security software. According to the company, that feature can identify up to 60 types of agents employees bring into the workplace, including Claude Code, Codex and OpenClaw. Discovery identifies software already in use, while access restrictions govern what that software can do. Microsoft’s proposal places these checks outside the model itself and assigns organizations responsibility for the actions taken through their accounts and systems.[1]
People define policy and consequential decisions
Microsoft chief executive Satya Nadella has called for separating the supply of model intelligence from authority over its use. Security chief Hayete Gallot describes frontier models as unpredictable systems trained to accomplish tasks rather than apply organizational judgment. In the announced design, people establish policy, approval boundaries and delegation rules. Playbooks coordinate protective actions, while Guardian Agents investigate changes in risk and can recommend or apply containment. Microsoft says independent systems must ultimately determine and enforce permitted actions. These are the company’s stated security mechanisms and operating principles; the announcement does not provide an independent measurement of their effectiveness.[1], [2]
