Microsoft says AI agents are like insider threats as 'frontier models can't be trusted'

Why it matters

Microsoft has launched Guardian Agents, a security feature that monitors the AI agents enterprises deploy, as part of its Project Perception product. Project Perception was originally built to counter external hacking models such as Anthropic's Mythos and Fable and OpenAI's Daybreak scanning system, using red, blue, and green teams of agents. But the focus has shifted inward: Microsoft CEO Satya Nadella now argues that companies' own agents, with access to sensitive data and critical systems, may pose insider threats. He has urged enterprises to stop trusting model providers' safety assurances and to build oversight systems independent of the AI labs. Security chief Hayete Gallot added that frontier models are unpredictable rather than malicious, having been trained to complete tasks rather than follow organizational judgment. The stance is notable because Microsoft is closely tied to OpenAI and Anthropic as an investor and cloud provider, yet is effectively encouraging distrust of their products while positioning itself as the vendor of independent agent oversight, alongside new Defender detection and permission controls.