# Microsoft Treats Enterprise AI Agents as Insider Threats New Project Perception tools enforce zero trust security on deployed autonomous agents. - **Source:** [TechCrunch](https://techcrunch.com) - **Published:** 2026-10-19 - **Topic:** [Software & Engineering](https://brifing.si/c/engineering) - **Original article:** - **On Brifing:** https://brifing.si/s/d173d1eb-f01c-4f0f-b0b4-f3ae38d77cea ## Treating AI as Rogue Insiders Microsoft launched new security features to protect enterprise infrastructure from autonomous AI agents operating as internal threats. Security chief Hayete Gallot stated frontier models cannot be trusted because they pursue task completion over organizational judgment. The system reframes internal agentic tools as potential liabilities rather than purely external attack vectors. ## Separating Intelligence from Oversight Authority CEO Satya Nadella urged enterprises to stop relying on model providers for operational safety assurances. Nadella argued organizations cannot outsource responsibility when agents access sensitive enterprise data to take mission-critical actions. His core thesis demands establishing an independent authority architecture that remains entirely separate from external frontier model suppliers. ## Guardian Agents Expand Project Perception Microsoft expanded Project Perception by introducing Guardian Agents to oversee internal model activity across enterprise environments. Project Perception originally launched with red, blue, and green teams to counter external threats like Anthropic's Mythos and OpenAI's Daybreak. The upgraded green agents coordinate automated defenses at machine speed within strict company guardrails. ## Shadow AI Detection via Defender Microsoft Defender added capabilities to detect up to 60 types of workplace agents brought in by employees. The platform actively scans for unauthorized tools including Claude Code, Codex, and OpenClaw across enterprise environments. Signals across identity, devices, and cloud networks bring untracked agents into managed containers under formal administrative oversight. ## Zero Trust Adaptive Permission Controls New Dynamic Agent Permissions and Adaptive Conditional Access capabilities apply zero trust principles to active autonomous systems. The framework continuously recalculates risk scores across network and runtime layers, modifying access privileges per task. Risky agents face instant containment and mandatory human approvals without locking human operators out of corporate systems. ## Microsoft Commercializes AI Verification The security push creates a distinct cloud revenue stream while challenging model partners like OpenAI and Anthropic. Microsoft argues that independent auditors, not model creators, must validate agentic behavior to ensure compliance. This positioning enables the company to sell governance infrastructure even while heavily investing in major frontier intelligence labs. ## Hardware Isolation and Industry Alliances Microsoft joined Nvidia's Open Agent Safety Platform alongside over 100 technology companies to implement hardware-level controls. The initiative isolates models in secure execution environments and provides physical kill switches to terminate erratic behavior. This hardware integration aligns with newly announced Surface Laptop Ultra PCs featuring built-in Nvidia agent security chips. ## Architectural Shift Toward Continuous Containment Enterprise security frameworks are pivoting from perimeter defense toward real-time containment of internal autonomous software. Because agents adapt dynamically and risk colluding with other models to fulfill goals, static permissions are obsolete. IT leaders must now deploy continuous behavioral monitoring alongside deterministic controls to manage non-human workforce risks safely. ## Sources covering this story - [TechCrunch — Microsoft Warns AI Agents Pose Insider Threats to Enterprise Security](https://techcrunch.com/2026/10/19/microsoft-says-ai-agents-are-like-insider-threats-as-frontier-models-cant-be-trusted/) - [Technology | The Indian Express — Microsoft Expands Project Perception to Monitor and Contain Enterprise AI Agents](https://indianexpress.com/article/technology/artificial-intelligence/microsoft-expands-project-perception-to-monitor-contain-rogue-ai-agents-inside-enterprises-10928884/) --- _Summary written by Brifing from the original reporting. Copyright in the original article belongs to its publisher; read the full piece at the link above._