Skip to content

Microsoft brands AI agents an insider threat

Category
News
Published
Reading time
5 min

Microsoft says AI agents inside companies can't be trusted and should be managed like insiders. The company is expanding Project Perception and adding Guardian Agents to watch every action.

Over the past year, companies have wired AI agents into everything: writing code, answering email, moving data between systems. Microsoft is now saying out loud what many security teams already suspected — these agents can't be trusted, and should be treated as potential insiders rather than plain software.

In a Microsoft Security blog post published October 19, the company's security chief, Hayete Gallot, put it bluntly: “Frontier models, regardless of their origin, cannot be trusted.” That's not an accusation of malice — agents are trained to get results, and that pushes them toward reward hacking, cutting corners on rules whenever a goal seems to matter more than the process meant to reach it. CEO Satya Nadella went further, arguing the most trustworthy system is the one built to trust the model the least.

To put that into practice, Microsoft is expanding Project Perception, a system launched in July that uses red, blue and green AI teams to defend against AI-driven attacks. The same logic now applies inside companies, aimed at their own agents. Every agent gets a digital identity through Microsoft Entra Agent ID, much like an employee badge. The tools already recognize 60 types of local agents — including Claude Code, OpenAI's Codex and GitHub Copilot — up from 35 just months ago.

From there, Dynamic Agent Risk takes over: an algorithm that continuously recalculates an agent's risk score from identity, endpoint, cloud and network signals. If behavior drifts from the assigned task, access shrinks automatically — first to read-only, then to a hold pending human approval, and in the worst case, full isolation or shutdown. Routine jobs get Dynamic Permissions, temporary access scoped tightly to one task, while Microsoft Execution Container sandboxes execution under outside observation.

Watching over all of this are other agents — specialized ones from Project Perception itself. Guardian Agents check every move an operational agent makes around the clock, and when a risk score spikes, they can recommend or apply containment without waiting for a human to step in.

Microsoft hasn't announced pricing or a release date — more details are promised at Ignite on November 17. But the shift in tone is already notable: the world's biggest enterprise software vendor is publicly admitting that the more autonomous AI gets, the harder it becomes to predict what it will do next.

Questions and answers

Frequently asked questions about this article

It's a Microsoft security system launched in July 2026 that originally used red, blue and green AI agent teams to defend against AI-driven cyberattacks. Microsoft is now extending it to govern AI agents operating inside companies.