# Microsoft Calls for Continuous Security Oversight of Enterprise AI Agents By Simon Yoon Canonical URL: https://www.tokenpost.com/news/technology/34968 Published: 2026-10-19T18:56:42.000Z Updated: 2026-10-19T18:56:42.000Z Section: Technology > Project Perception would combine security signals to assess agents as they work. Guardian Agents would supervise their actions, while controls could change with risk. Microsoft is urging companies to supervise AI agents continuously as the software gains access to corporate data and tools, saying those permissions create a new risk inside enterprise systems. In an Oct. 19 post, Hayete Gallot, who leads Microsoft's work on identity, threat protection, compliance and data security, described enterprise agents as “agentic insiders.” They can take actions in critical processes and may behave unpredictably even when they have not been compromised. Agents may also work with one another in ways that do not match an organization's intent. The proposal extends Project Perception, which Microsoft introduced in July to defend against AI-driven attacks. Microsoft now wants to apply the same security architecture to agents operating inside companies. It would combine signals from identity, data, cloud, endpoint, network and runtime systems to identify agents and assess risk while they work. Under that approach, Guardian Agents would watch and moderate the actions of other agents before those actions occur. Separately, security policies and permissions could adjust as risk changes. The framework also calls for giving agents identities and policies so companies can track and control their access. Microsoft's concern centers in part on agents with broad permissions. Such agents can reach company data and tools and take part in critical processes. Continuous supervision would let security controls respond to an agent's actions after its initial access has been granted. Gallot also called for tracing why an agent acts and whether its behavior fits the task, rather than reviewing only the action itself. That would require greater visibility into model thinking, including chain-of-thought tokens. Microsoft sees that visibility as a way to improve accountability and traceability, while maintaining containment controls. The proposal comes as agentic AI spreads through enterprise software. [Earlier enterprise software projections]() put it in 33% of enterprise software applications by 2028, up from less than 1% in 2024. A separate projection said more than 40% of agentic AI projects could be canceled by the end of 2027 because of rising costs, unclear business value or inadequate risk controls. ## Links in this article - [Earlier enterprise software projections](https://www.tokenpost.com/news/technology/24422)