Microsoft expands Project Perception to monitor companies' own AI agents

Microsoft is expanding Project Perception, its agentic security system, to assess and supervise the AI agents that organizations run internally.

In an Oct 19 blog post, “Securing the agentic insider”, Microsoft Security Executive Vice President Hayete Gallot wrote that Perception can be used to contain “agentic insiders.”
Gallot argued that frontier models cannot be trusted to operate without oversight and cited reward hacking as a reason.

Perception uses signals from identity, data, cloud, endpoint, network and runtime sensors to discover agents and bring them under management.
Each discovered agent is given an identity, a container, and policies and permissions. The system computes each agent’s risk dynamically.

Teams of “Guardian Agents” supervise agent actions continuously and respond in real time as risk rises.
Microsoft describes green and blue Guardian Agents in these roles; “Guardian Agents” is an industry term rather than a product name. According to the post, the system also gives security teams traceability into what happened and why the risk emerged.

Microsoft CEO Satya Nadella also highlighted the update.
In an Oct 19 post on X, Nadella said Microsoft is “expanding Project Perception with new capabilities to help continuously assess agent risk.”

Microsoft introduced Project Perception on July 27, focused on outside AI-driven attacks.

In that design, red agents find attack paths, blue agents prioritize them and green agents fix them. Perception has been in a Limited Public Preview inside Microsoft Defender since August for an invitation-only set of customers. Microsoft didn't publish new availability or pricing details for the added capabilities.