Microsoft Introduces New Security Tools to Address Risks From AI Agents

 

Microsoft has introduced new security features designed to protect enterprises from the growing risks posed by AI agents. The move reflects a broader warning from the company: AI agents should not automatically be trusted, even when they come from established technology providers.

 

One of the key additions is Guardian Agents, a feature designed to monitor the behavior of AI agents deployed within organizations. It builds on Microsoft’s Project Perception, launched in June to help enterprises identify, assess, and address security vulnerabilities using AI-powered systems.

 

Initially, Project Perception included three types of agents: red-team agents that simulate attacks, blue-team agents that detect threats and evaluate vulnerabilities, and green-team agents that work to fix security weaknesses.

 

However, Microsoft is expanding its security focus beyond external attackers. The company now warns that AI agents operating inside organizations could themselves become security risks, much like insider threats caused by employees who misuse their access.

 

Microsoft CEO Satya Nadella has argued that enterprises need independent systems to oversee AI agents, rather than relying solely on the assurances of the companies that develop them. As AI agents gain access to sensitive information, business applications, and critical systems, organizations must retain control over what those agents can do.

 

Microsoft security executive Hayete Gallot reinforced this concern, stating that frontier AI models cannot be trusted completely because their behavior can be unpredictable. These models are trained to accomplish tasks, but they may not consistently understand or follow an organization's policies, expectations, and judgment standards.

 

Microsoft is responding with several additional security measures. Its Microsoft Defender product can detect up to 60 types of AI agents used in the workplace, including Claude Code, Codex, and OpenClaw. The company is also introducing controls that restrict agent access to data and systems, adjust permissions according to tasks, and require human approval when potentially risky actions are identified.

 

The strategy also highlights a growing debate in the AI industry. Microsoft has major business relationships with OpenAI and Anthropic, operates its own AI development efforts, and provides cloud infrastructure for AI workloads. At the same time, it is positioning itself as a provider of independent security and governance tools for AI agents.

 

The underlying principle is straightforward: the companies supplying AI intelligence should not be the only ones responsible for controlling how that intelligence acts. Independent monitoring, restricted permissions, and human oversight can help enterprises reduce the risks associated with increasingly autonomous systems.

 

This approach is gaining wider industry support. Microsoft has joined Nvidia's Open Agent Safety Platform, an initiative involving more than 100 technology companies that aims to isolate AI agents in secure environments and monitor or stop unsafe behavior.

 

As enterprises adopt AI agents for increasingly important business functions, security is becoming more than a question of preventing external cyberattacks. Organizations must also ensure that the agents operating within their own systems remain accountable, constrained, and under human control.