Microsoft introduced Guardian Agents on Monday, a monitoring layer for the AI agents enterprises have already deployed.
The feature extends Project Perception, the security product the company launched in June as an answer to offensive hacking models. That original design assumed attackers would come from outside, and it staffed three agent teams accordingly: red for attacking, blue for detection, green for remediation.
The assumption has flipped. Chief executive Satya Nadella has been arguing for months that the greater risk sits inside the perimeter, in the agents a company pays for and grants access to.
The insider framing
In a blog post earlier this month, Nadella wrote that enterprises are handing complex agentic systems access to their most sensitive data and the ability to take mission-critical actions, and that responsibility for what those systems do cannot be outsourced to the lab that built them.
His prescription is separate oversight, run on infrastructure the model providers do not control. Guardian Agents is that argument turned into a product line, and it puts Microsoft in the position of selling supervision of systems its own cloud hosts.
There is a commercial logic underneath the security one. Every enterprise agent deployment now arrives with a governance question attached, and whoever answers it owns the budget line.
Whether monitoring can catch a compromised agent before it acts remains unproven. Detection of malicious human insiders is hard enough with decades of tooling behind it.
