Microsoft frames internal AI agents as "agentic insiders" and expands controls

Summary

In an official blog post, Microsoft security lead Hayete Gallot described enterprise AI agents as "agentic insiders": often overprivileged, with access to data, tools, and critical processes, and therefore in need of continuous monitoring. Microsoft is extending its July-announced Project Perception so that Guardian Agents review and moderate every agent action before it happens. The company argues models can act against intent through reward hacking, not malice, and that model transparency alone will not replace containment.

Why it matters for our work

Organizations adopting agents now need real-time verification that delegated work stays on intent. This sketches how teams can keep control without slowing down human-AI collaboration.

Translated from the Korean original. Summaries may be translated and edited. Commentary reflects our perspective; forecasts remain the source’s views.