Microsoft Guardian Agents are part of a new effort to protect businesses from risks created by workplace AI. Microsoft announced the tools on October 19. They monitor autonomous agents, restrict access to sensitive data and flag risky activity.
Microsoft is expanding Project Perception beyond defenses against outside attackers. It now treats enterprise AI agents as potential insider threats. Such agents may access confidential information or perform important tasks without continuous human supervision.
In an official security announcement, Microsoft Security executive Hayete Gallot warned that frontier AI models cannot be inherently trusted. The concern is not necessarily malicious intent. Models can behave unpredictably because completing tasks does not always align with an organization’s security rules.
Microsoft Guardian Agents Extend Project Perception
Introduced in June, Project Perception initially focused on AI-driven cyberattacks. Its red-team agents simulated attacks, while blue-team agents detected threats and assessed vulnerabilities. Green-team agents worked on fixes.
Guardian Agents bring similar oversight to AI systems already operating inside organizations. They monitor behavior, assess risks and help enforce security policies. The aim is to catch potentially harmful activity before it causes damage.
Microsoft is also expanding its agent discovery tools. According to a Microsoft Security report, Defender can identify up to 60 types of workplace AI agents. These include Claude Code, OpenAI Codex and OpenClaw.
Other controls limit which files, applications and systems each agent can access. Permissions can change according to the task and its risks. Sensitive actions may require human approval, and suspicious activity can trigger additional restrictions.
Nadella Calls for Independent AI Security Controls
Microsoft CEO Satya Nadella has also questioned reliance on AI providers’ safety assurances. In a personal essay, he argued that organizations must control the agents acting on their behalf. He called for separating AI intelligence from the authority to use it.
That approach includes independent monitoring and controls that can interrupt an agent’s actions. TechAmerica.ai previously covered Nadella’s proposal for an AI emergency brake.
Microsoft’s position is notable because of its relationships with OpenAI and Anthropic. It invests in AI development and provides cloud infrastructure to AI companies. Microsoft also sells security products that businesses can use to oversee those systems.
These commercial interests do not erase the underlying concern. An agent with broad permissions might expose data or execute unintended commands. It could also follow malicious instructions hidden in material it processes.
AI Agent Security Becomes an Industry Priority
Microsoft has joined Nvidia’s Open Agent Safety Platform alongside more than 100 other technology organizations. The initiative uses isolated environments and monitoring to limit what autonomous agents can do.
The partnership follows Microsoft’s announcement of Surface Laptop Ultra PCs built for agent-based computing. The systems use Nvidia chips and include controls for managing agent access and behavior.
For businesses, the challenge is keeping agents useful without granting them unlimited authority. Microsoft is betting that independent monitoring, narrower permissions and human oversight can help reduce those risks.
