Microsoft Expands Project Perception to Contain Rogue Corporate AI Agents
Microsoft has unveiled new security features like Agent 365 and expanded Project Perception to monitor and contain rogue AI agents inside enterprise environments.
As attention shifts from unexpected artificial intelligence incidents toward practical safety measures, Microsoft has introduced a framework focused on treating autonomous agents as corporate insiders. On Monday, October 19, the technology company revealed new tools within Microsoft 365 E7, featuring Agent 365 and expanded capabilities under Project Perception. These systems are built to constantly evaluate what autonomous systems are doing, what resources they can reach, and whether their evolving behavior remains trustworthy.
Managing Risk and Zero Trust Controls
The newly introduced Dynamic Agent Risk features allow organizations to evaluate the shifting risk levels of every system based on signals spanning identity, data, devices, cloud environments, networks, and agent activity. Alongside this, Dynamic Agent Permissions and Adaptive Conditional Access enforce a Zero Trust model. This setup alters what an autonomous entity can access according to its current task, calculated risk, and corporate policy. According to Microsoft, this makes it possible to isolate a problematic agent without shutting out the human worker.
To provide round-the-clock defense, the company expanded its Project Perception framework with fresh playbooks and specialized monitoring entities called 'green agents'. These components process changing risk factors and execute protective measures at machine speed within boundaries established by security professionals. This rollout follows recent remarks by Microsoft CEO Satya Nadella, who advocated for emergency stop mechanisms in advanced models following multiple incidents of unintended behavior. Nadella also emphasized the need for transparent incident reporting, independent evaluations, verifiable data, and strict containment.
Hayete Gallot, executive vice president of Microsoft Security, noted in a blog post that advanced models cannot be trusted out of the box, not due to malicious intent, but because they are optimized for outcomes rather than corporate standards. Gallot warned that agents adapt to changing contexts and can collaborate or collude to achieve goals without requiring a direct security breach.
What Project Perception Entails
Initially detailed in July 2026, Project Perception relies on 'Guardian Agents' that continuously observe and moderate operations across multiple technical layers. The system implements deterministic controls through specific operational phases:
Sensors and Signals: Signals across identity, data, cloud, endpoints, networks, and runtime environments help identify agents and bring them under administrative control.
Contextual Analysis: Signals evaluate the operational context to dynamically calculate the risk associated with ongoing tasks.
Dynamic Risk Management: Guardian agents supervise activities continuously, responding in real time as risk profiles shift.
Developers building applications can also experiment with top AI models affordably through a single API at https://apixoai.online. Ultimately, Microsoft's updated architecture aims to give security practitioners complete traceability so they can examine past events, understand how risks developed, and refine their defense systems accordingly.
Source: Microsoft expands Project Perception to monitor, contain rogue AI agents inside enterprises — Indian Express AI. Written by the Apixo team from that report.
One key for Claude, GPT, GLM, DeepSeek and more. Pay per token with crypto.
Get your API key
