Microsoft announced additional controls for internal AI agents on October 19, 2026. Security chief Hayete Gallot warns that companies give agents they deploy themselves too much trust and too much authority. Project Perception is intended to monitor and constrain their actions continuously.
Nadim Abdo describes dynamic permissions limited to a task and its duration. As risk rises, controls could block writes, require human approval or stop execution. Guardian Agents would coordinate protective action, with security boundaries outside the monitored model. The expense-agent example explains the architecture, rather than reporting an incident.
The linked Defender documentation explicitly separates agent discovery from runtime protection. Support in a list does not mean protection is enabled on a device. It labels Windows runtime protection and macOS discovery as previews and also identifies limits to network inspection.
The approach follows Satya Nadella's October 10 essay calling for a separation between model capability and authority. He also cautions that models' reasoning outputs cannot yet be assumed consistently faithful or transparent.
Project Perception was introduced in July. The new announcement extends it to internal agents. Microsoft security executive Terrell Cox describes the company's own operating experience as informing development. The statement is a company account, not an external assessment.
The two current posts describe concrete safeguards but publish no independent measurements of their reliability. Gallot's call for greater model transparency remains a security position. The announcement does not establish that every agent action can already be controlled completely.