Microsoft Flags AI Agents as 'Insider Threats'

Microsoft is shifting its stance on AI agents, warning enterprises, as TechCrunch first reported, that the tools they pay for could become "insider threats" similar to rogue employees. Microsoft has introduced Guardian Agents, a monitoring feature integrated into Project Perception. While Project Perception originally focused on external threats—specifically targeting models like OpenAI's Daybreak and Anthropic's Fable and Mythos—the company is now focusing on the risks posed by deployed agents.

Microsoft CEO Satya Nadella stated in a private blog that enterprises must stop relying on safety assurances from AI model providers. Nadella argued that companies cannot outsource responsibility for the actions of intelligence systems, asserting that the "supply of intelligence" must be separated from the "authority over it."

Hayete Gallot, Microsoft's security business chief, told TechCrunch that frontier models cannot be trusted regardless of their origin. Gallot noted that these models are unpredictable because they are trained for task completion rather than organizational judgment, yet are often granted extensive access to sensitive data and tools.

To capitalize on this risk, Microsoft is expanding its security ecosystem. Beyond Guardian Agents, the company added a detection feature to Microsoft Defender capable of identifying 60 agent types, including OpenClaw, Codex, and Claude Code. New features also allow for dynamic permission adjustments and mandatory human approval for risky decisions. Additionally, Microsoft is partnering with Nvidia on the Open Agent Safety Platform—supported by over 100 companies—to implement hardware-based "kill switches" for agents. Microsoft has also released Surface Laptop Ultra PCs featuring built-in agent security controls powered by Nvidia chips.