Microsoft warns AI agents pose insider risks, launches Guardian Agents

US technology company Microsoft issued a warning on October 19, 2026, stating that the widespread deployment of AI agents within enterprises could evolve into a new cybersecurity risk. The company emphasized that frontier models, regardless of their supplier, "cannot be trusted," as they are trained to pursue results and may deviate from organizational norms like an "insider threat," compromising sensitive data and critical processes without external hacking.

To address this risk, Microsoft argued that enterprises cannot leave security responsibility entirely to model suppliers but must establish control mechanisms independent of the models themselves. The company announced an expansion of its Project Perception security architecture, adding new capabilities such as dynamic permissions, dynamic agent risk assessment, and dynamic conditional access. These mechanisms can adjust available permissions in real-time based on changes in agent risk, restricting access scopes when necessary or requiring human approval for high-risk actions.

Additionally, Microsoft introduced so-called Guardian Agents, serving as a protective layer to continuously supervise other AI agents. Microsoft stated that these guardian agents will monitor the behavior of every AI agent around the clock, intervening immediately when risks escalate to help enterprises bring non-deterministic AI behaviors under controlled scope. Microsoft CEO Satya Nadella confirmed this move on social media, describing it as putting the principle of treating models as insider risks into practice.