Microsoft Launches Security Controls to Treat AI Agents as Insider Risks

Microsoft introduced new security controls designed to treat AI agents as potential insider risks. Announced by Nadim Abdo, corporate vice president of identity and network access, the tools extend Project Perception to limit agent access and automatically tighten restrictions if an agent's behavior changes.

The system operates on a zero-trust principle where agents start with no authority and receive only the permissions necessary for specific approved tasks. Key features include the Microsoft Entra Agent ID for identity tracking, dynamic permissions, and a risk assessment system that monitors for behavioral drift. To prevent agents from bypassing security, the company is implementing sandboxed runtimes and ensuring security controls remain independent of the AI models.

Microsoft is acting as Customer Zero, deploying these tools across hundreds of thousands of its own agents to ensure scalable adoption. The company aims to maintain security even when AI models are unpredictable, ensuring that the infrastructure protecting the system is separate from the models themselves.