What happened
At Microsoft Build 2026 on 2 June 2026, Microsoft announced the Microsoft Execution Container (MXC) SDK — an early-preview policy-driven runtime isolation layer for local AI agent workloads on Windows — and extended the Agent 365 control plane to cover local agents and 'claws' (persistent autonomous processes). Capabilities include agent observability for 20+ local agent runtimes across managed Windows and macOS devices, tool-level policy enforcement, and integration with Defender, Entra, Intune, and Purview. Agent 365 integration with MXC is planned for July preview, and Windows 365 for Agents (cloud PCs for computer-using agents) is generally available. Microsoft also announced MDASH improvements — its multi-agent vulnerability research system used to discover AI-stack vulnerabilities.
Why it matters
Local agents running on endpoints inherit user credentials and run with minimal central oversight — a risk profile that existing endpoint security tools were not designed to address. Microsoft's platform approach to agent containment and observability raises the baseline control expectation for enterprise agent deployments and will influence what CISOs require from any agent vendor.
Applicability
Enterprises deploying GitHub Copilot, Windows Copilot, or any local coding/agentic AI should evaluate MXC and Agent 365 as the foundational layer for endpoint agent governance; security architects should map agent identity and containment requirements against these controls before production deployments.