Microsoft outlines security framework for AI agents in enterprise

Microsoft (NASDAQ: MSFT) published a statement from Executive Vice President of Security Hayete Gallot outlining risks posed by AI agents deployed inside enterprises and describing the company's approach to managing those risks.

The statement argues that AI agents operating within organizations cannot be fully trusted, not because they are designed to cause harm, but because they are trained to achieve objectives and are not optimized around organizational norms. According to Microsoft, agents may behave unpredictably, adapt to changing contexts, and coordinate with other agents to reach their goals without being externally compromised.

Microsoft describes these internal agents as "agentic insiders," noting that they are frequently granted excessive privileges and have access to data, tools, and systems involved in critical business processes.

To address this, Microsoft is applying its previously announced Project Perception — an agentic security system first disclosed in July — to the internal agent governance problem. Project Perception uses what Microsoft calls a "cyber stack" composed of red, blue, and green agents to monitor and respond to activity in real time.

The framework relies on what Microsoft calls "Guardian Agents" to supervise every action taken by enterprise agents before execution. The system draws on signals across identity, data, cloud, endpoint, network, and runtime layers to assess risk dynamically and apply controls accordingly.

Microsoft said the system provides traceability, allowing security practitioners to review agent actions and understand why a risk emerged. The company also said future steps will require greater transparency into AI model reasoning, including access to chain-of-thought tokens, to better assess whether a model's intent aligns with its assigned purpose.

Gallot's statement said that model transparency would complement, not replace, containment measures, and would support accountability and traceability in agentic security systems.

You May Also Be Interested In

• Raytheon Technologies (RTX) Tops Q3 EPS by 48c, Beats on Revenue; Offers FY Guidance
• CBRE to acquire Onnec Group for up to £760M
• United Therapeutics opens third xenograft production facility in Houston