Microsoft releases open‑source toolkit to secure AI agents at runtime

Microsoft has introduced a new open‑source toolkit focused on securing AI agents at the runtime level, that is, while they are actually executing actions. The tool targets enterprise systems where autonomous or semi‑autonomous agents can initiate real‑world operations in business processes — from querying internal databases to interacting with external services. It allows organizations to define and enforce policies that constrain agent behavior and to monitor attempts to go beyond those boundaries.

In effect, Microsoft is offering a “governance by default” layer for enterprise AI: security and control policies are embedded directly into the infrastructure through which agent traffic flows, rather than being left to each individual application. This is particularly relevant given the growth of so‑called shadow AI — unsanctioned agent deployments by teams and employees outside formal IT oversight. The new toolkit is designed to close that gap by giving IT and security teams a single control plane for agent activity.

Observers say the arrival of such tooling marks a shift from simply showcasing what autonomous agents can do to managing how they operate in tightly regulated corporate environments. For developers and architects, it is a clear signal that agent safety and policy enforcement are becoming as mandatory a part of the stack as authentication, logging, or encryption.