Microsoft Unveils Security Tools Treating AI Agents as Insider Threats
Microsoft introduces new features in Microsoft 365 E7 and Project Perception to monitor AI agent risks via continuous assessment and dynamic permissions.
ImportanceMaterialEvidenceE3 inspectableWrite-upStandard
On October 19, 2026, Microsoft announced new security tools within its Microsoft 365 E7 suite and Project Perception system designed to treat AI agents as potential "insider threats." These features are currently in private preview, with broader availability planned for the Ignite conference in San Francisco this November.
The core mechanism involves continuously assessing the risk level of each agent based on signals such as the data it can reach and whether its actions are deviating from assigned tasks. When risk levels rise, the system dynamically adjusts permissions: it may revoke modification rights while retaining read access, require human approval for high-stakes steps, or shut down the agent entirely. Additionally, Microsoft has expanded its ability to detect AI agents running on employees' computers, increasing the number of detectable agent types from 25 to 60, including popular tools like Anthropic's Claude Code, OpenAI's Codex, and GitHub Copilot.
This move builds on Microsoft CEO Satya Nadella's recent argument that controls on AI models must come from outside the models themselves, rather than relying on inherent model trustworthiness. Microsoft is also introducing "Guardian Agents," AI security proxies that investigate why another agent's risk profile changed and recommend or execute containment steps. The announcement intensifies competition with rivals such as Google, Amazon Web Services, CrowdStrike, and Okta in the space of enterprise AI control.