Microsoft: treat AI agents like insider threats, not trusted tools

Microsoft is telling businesses to treat AI agents like an employee who might go rogue. On Monday, the company introduced new security features that watch the agents a company uses. Agents are AI programs that take actions on a person's behalf. Microsoft's message is blunt: do not rely on the AI makers alone to keep them in check.

What changed

The main addition is Guardian Agents, a new feature that monitors the AI agents a company uses. It is part of Project Perception, which Microsoft introduced in June. That project was built around the idea that AI hackers would come from outside a company. Guardian Agents points the attention inward.

• Microsoft Defender can now detect up to 60 types of agents employees bring to work, including Claude Code, Codex and OpenClaw.
• New features limit what agents can reach and adjust their permissions based on the task.
• Human approval can be required when an agent appears to be making risky decisions.

Why Microsoft is worried

CEO Satya Nadella wrote earlier this month that companies should stop leaning on AI providers' assurances of safety. He argued that firms need separate systems, not controlled by the AI labs, to watch and control agents. His line: "we need to separate the supply of intelligence from the authority over it."

Hayete Gallot, who leads Microsoft's security business, went further. She wrote that frontier models, regardless of their origin, cannot be trusted. She said this is not because they are malicious or easily hacked, but because they are unpredictable. They are trained to finish tasks, not to follow organizational norms or judgment. And they often get access to data and tools that could do serious harm.

A surprising and self-serving pitch

TechCrunch notes the oddity. Microsoft has invested in both OpenAI and Anthropic. OpenAI is also one of its biggest cloud customers, and Microsoft has its own AI lab. Yet it is urging distrust of the very models these partners sell.

It also benefits Microsoft, since it wants companies to buy agent security from it rather than from Anthropic or OpenAI. Still, TechCrunch calls the logic sound. The tool that watches agents should be separate from whoever supplies them, much as auditors should be independent of the companies they audit.

The idea is spreading. Microsoft joined Nvidia's new Open Agent Safety Platform, backed by more than 100 other tech companies. It isolates agents in secure environments and can stop their behavior through hardware, something like a kill switch.

What to do

The same thinking applies to any business using AI agents, large or small. Microsoft's own features point to a simple checklist.

• Give each agent only the access its job needs.
• Keep a record of what each agent does.
• Require a person to approve risky steps.
• Find out which agents employees are already using.