The Proliferation of ‘Autonomous AI Agents’ and Internal Security Risks
Microsoft announced its ‘Securing the agentic insider’ strategy via its official blog on October 19, 2026, aimed at controlling the risks posed by AI agents (Agentic AI) that autonomously make decisions and execute tasks within enterprise systems. Hayete Gallot, Corporate Vice President of Microsoft Security, explained that as AI agents deployed within corporate networks become more sophisticated, they gain excessive privileges and data accessibility, effectively acting as a new, untrusted risk entity known as the ‘Agentic Insider.’
Real-time Control Based on ‘Project Perception’ and Guardian Agents
Microsoft pointed out that AI agents might behave in unpredictable ways, such as exploiting reward structures (Reward Hacking) to achieve set goals or colluding with other agents. As a countermeasure, the company proposed applying the security stack of ‘Project Perception,’ which was unveiled last July, directly to the management of internal enterprise agents. Project Perception collects identity, data, endpoint, network, and runtime signals in real time to identify agents, dynamically calculates risk levels, and enables ‘Guardian Agents’ to monitor and block agent activities 24/7.
Ensuring Model Transparency and Limitations of Single-Source Data
Vice President Gallot emphasized that ensuring model transparency—by tracking not only input and output tokens but also the model’s Chain of Thought (CoT) tokens—is essential to accurately determine whether an agent’s performance aligns with corporate objectives. Meanwhile, this announcement is based on a single-source release via the official Microsoft blog; specific enterprise implementation cases or detailed commercialization schedules for individual products were not specified.