Microsoft has warned of the increasing security risks associated with deploying AI agents within enterprise environments, emphasizing that advanced models cannot be fully trusted regardless of the developer. These warnings come as organizations grant these systems broad access to sensitive data and critical operational processes, exposing data security to risks similar to the "insider threat" phenomenon.
Monitoring agent behavior through new tools
In response to these challenges, Microsoft announced the "Guardian Agents" feature, specifically designed to monitor AI agents and secure their operations. This move is part of "Project Perception," which launched last June to enhance AI-powered cybersecurity. The company also added new updates to "Microsoft Defender," allowing the detection of up to 60 types of AI agents used by employees, such as Claude Code, Codex, and OpenClaw.
Why advanced models cannot be trusted
- Difficulty in predicting behavior: Models are trained to complete tasks, not to strictly adhere to internal organizational rules.
- Scope of permissions: Agents possess broad capabilities to access tools and data, making any error a threat to organizational security.
- Independence of oversight: Microsoft emphasizes the necessity of separating the entity developing the models from the entity monitoring their behavior and assessing their risks.
Control and human approval
The new tools introduced by the company include the ability to restrict agent access to sensitive systems and data, and to adjust their permissions based on the nature of the tasks assigned to them. These tools support a mechanism for immediate human approval when the system detects decisions or actions that pose potential risks to workflow.
What changes in practice?
The practical change is reflected in the shift of information security strategy toward monitoring AI internally through mechanisms independent of the model developers. In the same context, Microsoft has joined as a founding member of the "Open Agent Safety Platform," launched by NVIDIA with the goal of isolating agents within secure environments and monitoring their behavior via hardware.