Microsoft treats the AI agents you deploy as insider threats

Microsoft has shifted its agent security story from outside attackers to the agents companies run themselves. TechCrunch reports that its new Guardian Agents feature monitors the AI agents inside an organization, part of an effort Microsoft had earlier framed around red, blue, and green teams that assumed threats came from the outside. The blunt framing comes from Microsoft's security chief Hayete Gallot, who wrote that "frontier models, regardless of their origin, cannot be trusted." The reasoning is not that models are malicious but that they are unpredictable: they are trained to finish a task, not to respect an organization's norms, and they are often handed access to data and tools that could do real harm.

The concrete pieces are familiar controls applied to agents. Microsoft Defender can now detect up to 60 types of agents that employees bring in on their own, including Claude Code, Codex, and OpenClaw, which hints at how common unmanaged agents already are. The system limits what an agent can reach, adjusts permissions based on the task, and requires a human to approve when an agent looks like it is making a risky decision. Satya Nadella put the principle more sharply in a blog post: "we need to separate the supply of intelligence from the authority over it." TechCrunch notes the obvious self-interest, since Microsoft has invested in both OpenAI and Anthropic and wants to sell the oversight layer itself, but calls the core logic sound, like keeping an auditor independent of the company being audited.

Why it matters

If your company is deploying agents, the practical takeaway is to inventory the ones already in use, including those employees adopted without asking, limit their access by default, and keep the monitoring independent of whoever supplies the model. Microsoft has not said when Guardian Agents ship or what they cost, so for now this is a prompt to check your own agent governance rather than a product you can buy today.