Microsoft Targets Rogue AI Agents With Dynamic Permissions, Risk Scoring

Microsoft is enhancing its Project Perception security system to address risks posed by AI agents deployed internally in organizations. Recognizing that AI agents can act with significant privileges and impact critical processes, Microsoft treats them as insiders capable of mistakes or compromise. The new features include Dynamic Permissions, which start agents with minimal access and grant permissions per task with expiration or reduction as tasks change. Dynamic Agent Risk assigns continuously updated risk scores based on agent activity and environmental signals, influencing access controls via Dynamic Conditional Access. This access can be scaled back, limited to read-only, or halted if risk is high. Playbooks define security responses, and Guardian Agents execute containment actions. The system supports discovery of 60 agent types, including popular AI coding assistants.

The controls aim to reduce the risk of unauthorized or harmful actions by AI agents within enterprise systems by limiting their access dynamically and requiring human oversight for high-impact actions. This reduces the potential damage from compromised or malfunctioning AI agents. There is no indication of active exploitation or inherent vulnerability; rather, this is a proactive security enhancement to manage AI agent risks.

Microsoft has implemented these controls as part of Project Perception, providing Dynamic Permissions, risk scoring, and conditional access to manage AI agent privileges. Organizations should adopt these features to enforce least privilege and continuous risk assessment for AI agents. Human review is required for high-impact actions, and containment measures are automated via Guardian Agents. No additional immediate action is indicated beyond deploying and configuring these new controls as recommended by Microsoft.