Microsoft on Monday unveiled new security tools to monitor and control autonomous AI agents. The tools address potential insider threat risks. Microsoft integrates these features into the high-tier Microsoft 365 E7 and the Project Perception security system.
The system continuously assesses each AI agent's risk level. This assessment relies on the agent's data access and actions.
If an agent's behavior deviates from its assigned task or becomes risky, the tools automatically restrict its permissions. They can also require human approval for actions or shut the agent down. The framework includes "Guardian Agents" within Project Perception to supervise other agents' actions.
Microsoft treats AI agents as "agentic insiders." These insiders, while not malicious, may behave unpredictably to achieve their goals. This initiative addresses growing concerns about unintended actions from autonomous AI systems.