Microsoft warns enterprises to treat AI agents like insider threats

Microsoft is urging companies to treat the AI agents they deploy as potential insider threats, arguing that organizations should not rely solely on model providers to supervise systems with access to sensitive data and critical tools.

The warning accompanies new enterprise security features, including Guardian Agents, a monitoring capability within Microsoft’s Project Perception. Microsoft previously described that project as a system with red-team agents that attack, blue-team agents that identify vulnerabilities and green-team agents that help fix them. The newer approach focuses on agents operating inside the organization.

According to TechCrunch, Microsoft security executive Hayete Gallot said frontier models cannot be trusted—not because they are necessarily malicious or easily compromised, but because their behavior can be unpredictable. Models are trained to complete tasks, she said, rather than reliably follow an organization’s norms or exercise human judgment.

Microsoft CEO Satya Nadella has made a similar case, writing that companies need to separate the providers that supply AI intelligence from the systems that control its authority. In practical terms, that means using independent controls to monitor agents, limit their permissions and require human approval when behavior appears risky.

Microsoft also added agent detection to Defender, which the company says can identify as many as 60 types of workplace agents, including Claude Code and Codex. Other controls can restrict access to data and systems and adjust permissions according to the task.

The strategy also creates a business opportunity for Microsoft, which sells cloud and security products while maintaining close commercial relationships with AI providers including OpenAI and Anthropic. TechCrunch noted that Microsoft’s argument could therefore serve both a security rationale and an effort to position its own products as the oversight layer for enterprise AI.

The broader idea is gaining industry support. Microsoft is among the members of Nvidia’s Open Agent Safety Platform, which is designed to isolate agents, monitor their behavior and stop them when necessary. The shift reflects a growing concern that agent security cannot be left entirely to the companies that build the underlying models.