Microsoft executive Charlie Bell warned that enterprises should not blindly trust AI agents, as these systems can make errors or be manipulated, leading to internal security risks. The company introduced new safety measures to monitor agent behavior and enforce strict permission boundaries.
The core of Microsoft's strategy is to decouple AI decision-making from security enforcement. Independent systems, rather than the AI models themselves, will now manage access rights and operational limits to ensure that agents cannot exceed their authorized scope.
Microsoft is expanding its Project Perception system, which includes dynamic permission management and risk assessment. These tools allow enterprises to assign unique identities to agents, grant minimal necessary access, and automatically revoke permissions once tasks are completed.
The company is also deploying Guardian Agents, which act as supervisors to monitor other AI agents in real-time. These systems can investigate suspicious activity, restrict operations, or isolate agents if they deviate from their intended tasks or attempt unauthorized actions.
Microsoft plans to integrate these security features into products like Microsoft 365 to help organizations track and manage AI agents. The platform will support the identification of up to 60 types of local AI agents, including tools from Anthropic, OpenAI, and GitHub.
