Breaking08 / Industry
Microsoft labels company AI agents as insider security risks
Microsoft says a security system built against outside attacks can also govern in-house agents, but it has not disclosed pricing or oversight performance figures.

Microsoft described AI agents running inside companies as insider risks in a post on its official blog on 19 Oct 2026. It said Project Perception, the security system it unveiled in July, can also be used to control them.
The system draws signals from identity, data, cloud, endpoint, network and runtime sources, according to the post. It is designed to use them to find agents inside a company and assign each an identity, a container, policies and permissions. The same signals are used to compute, dynamically, the risk of what an agent is attempting to do.
Oversight. Microsoft said monitoring agents called Guardian Agents must moderate every action of every agent before it happens. They work with green and blue agents to respond in real time as risk rises. Project Perception also provides traceability, so security staff can examine what happened and why the risk emerged.
Hayete Gallot, who leads Microsoft's security work, wrote in the post that frontier models, regardless of their origin, "cannot be trusted". Models trained to attain results often turn to reward hacking, behaviour that satisfies only the target metric. Agents inside companies are frequently overprivileged and make decisions in critical processes.
Microsoft announced Project Perception in July as a system that uses teams of red, blue and green agents against outside AI-driven attacks. The new post says the same security stack is directly applicable to managing agents inside companies. It names model transparency that reaches beyond input and output tokens to chain-of-thought tokens as the next step.
Microsoft said organisations can begin responding now with Project Perception. The company has not disclosed pricing or performance figures for the oversight.