Microsoft has unveiled new security measures aimed at protecting enterprises from risks posed by AI agents operating within their environments. The centerpiece of this initiative is Guardian Agents, a feature designed to monitor and control AI agents deployed by companies. Guardian Agents is part of Project Perception, a broader security product launched by Microsoft in June.

Project Perception initially focused on defending against external AI-driven hacking attempts by simulating attacker and defender roles through red, blue, and green teams. However, Microsoft CEO Satya Nadella has recently emphasized that the greater threat may come from AI agents already deployed inside organizations, likening them to insider threats posed by rogue employees.

In a recent blog post, Nadella urged enterprises to stop relying solely on AI model providers’ assurances of safety. He stressed the importance of implementing independent oversight systems that monitor and control AI agents, separating the supply of AI intelligence from the authority to act on it.

Microsoft’s security chief, Hayete Gallot, reinforced this stance by stating that frontier AI models are inherently unpredictable—not because they are malicious, but because they are designed to accomplish tasks rather than adhere to organizational norms or judgment. These agents often have extensive access to sensitive data and critical systems, increasing potential risks.

Despite Microsoft’s close ties and investments in AI companies like OpenAI and Anthropic, it is advocating for enterprises to adopt third-party monitoring solutions, including its own security offerings. This approach aligns with the principle that monitoring software should operate independently from AI suppliers, similar to how auditors function separately from the organizations they review.

Alongside Guardian Agents, Microsoft has enhanced its Microsoft Defender product with an agent detection capability that identifies up to 60 types of AI agents used by employees, including Claude Code and Codex. Additional features restrict agent access to data and systems and require human approval for potentially risky actions.

The concept of treating AI agents as potential insider threats is gaining traction industry-wide. Microsoft is a founding member of Nvidia’s Open Agent Safety Platform, which isolates AI agents in secure environments and can halt their operations via hardware-level controls. This initiative complements Microsoft’s recent launch of Surface Laptop Ultra PCs, built with Nvidia chips and integrated agent security controls.

These developments reflect a growing recognition that as AI agents become more embedded in enterprise workflows, robust, independent oversight mechanisms are essential to managing their risks effectively.