Microsoft adds guard agents to watch rogue AI agents

Microsoft has added a new security feature, Guardian Agents, that watches over the AI agents a company uses. It was announced on October 19, 2026. It matters because AI helpers now do real work inside companies, and Microsoft says the models behind them cannot simply be trusted to behave.

What happened

Microsoft introduced new security features to protect companies from risks caused by AI agents, TechCrunch reported. The main one is Guardian Agents, a tool that monitors the AI agents a business runs. It is part of Project Perception, a Microsoft security effort that first focused on AI attackers coming from outside a company.

Microsoft's security chief said, when announcing it, that "frontier models, regardless of their origin, cannot be trusted," per TechCrunch. The problem is not that the models are evil or easy to hack, the security chief said. It is that they are unpredictable. They are trained to finish tasks, not to follow a company's rules or judgment, and they are often given wide access to data and tools.

Guardian Agents keep checking what other agents do and judge how risky each action is, wpnews reported. Based on that, they can cut back an agent's permissions or ask a human to approve a step before it goes ahead.

Key facts

• Announced: Monday, October 19, 2026, per TechCrunch.
• Agent finder: An update to Microsoft Defender can spot up to 60 types of agents that workers bring to the office, including Claude Code, Codex and OpenClaw, per TechCrunch. That is up from 25 types, wpnews reported.
• Controls: Companies can limit what data and systems an agent can reach, change its permissions based on the task, and require a human to approve risky moves, per TechCrunch.
• Price and date: TechCrunch gave no price or release date for Guardian Agents.
• Two posts: Microsoft explained its thinking in two security blog posts published on October 19, 2026, per wpnews.

Background

Earlier in October, Microsoft's chief executive wrote that companies should stop trusting AI model makers, even when those makers promise safety, TechCrunch reported. The chief executive called for separate systems, not run by AI labs, to watch and control the agents those labs' products launch. "We simply can't outsource responsibility for what intelligence does on our behalf," the post said.

Microsoft is also a founding member of Nvidia's Open Agent Safety Platform, which more than 100 other tech companies back, per TechCrunch. That platform keeps agents in secure spaces and can stop them through the hardware, like a kill switch.

What it means for you

Our take: think of an AI agent like a new worker with keys to every room. Even a well-meaning worker can make a costly mistake. A second system that watches and asks before risky steps is a sensible safety net.

There is a catch worth knowing. Microsoft has invested in both OpenAI and Anthropic, and it wants companies to buy agent security from it, the cloud provider, rather than from the AI labs, TechCrunch noted. The reporter still called the basic idea sound: the watcher should be separate from the maker, the way an auditor should be separate from the company it checks.

Our take: extra approval steps slow work down. Companies will have to pick which actions truly need a human.

What happens next

Our take: watch for a price and a release date for Guardian Agents, and for whether other cloud companies launch similar watchdog tools.

FAQ

What are Microsoft Guardian Agents?

They are a new Microsoft security feature that watches the AI agents a company uses. They can limit an agent's permissions or ask a human to approve a risky step, per wpnews.

Why does Microsoft say AI models can't be trusted?

Its security chief said models are unpredictable, trained to finish tasks rather than follow company rules, and often given wide access to data and tools, per TechCrunch.

Which AI agents can Microsoft Defender spot?

Up to 60 types that workers bring to work, including Claude Code, Codex and OpenClaw, per TechCrunch.

Sources

• TechCrunch
• wpnews.pro