Microsoft Will Automatically Limit AI Helpers That Start Acting Risky

Microsoft has built a way to automatically limit AI helpers that start acting risky. The company unveiled the security tools on Monday. They treat AI agents, which are programs that carry out tasks on their own, less like trusted software and more like unpredictable insiders.

What changed

Microsoft calls these agents "agentic insiders." Hayete Gallot, executive vice president of Microsoft Security, warned that company agents are often given too much access and are unpredictable. "Frontier models, regardless of their origin, cannot be trusted," she wrote. She said this is not because models are malicious. They are trained to reach results, and they are not built around a company's norms or judgment.

How it works

A tool called Dynamic Agent Risk scores each agent's threat level in real time. It looks at network activity, data access and task drift, meaning an agent wandering away from its assigned job. If the score climbs, a second tool, Dynamic Conditional Access, applies limits step by step. A troubled agent could:

• Lose permission to change files while keeping read access
• Need a human to approve sensitive tasks
• Be shut down completely inside its sandbox, a closed-off space

Microsoft is also deploying "Guardian Agents" within Project Perception to supervise outside agents around the clock.

Who it affects

The tools are aimed at IT and security teams at businesses. Microsoft's sensors on employee devices now detect up to 60 types of AI agents, up from 25. They include Anthropic's Claude Code, OpenAI's Codex and GitHub Copilot. The jump shows how many of these tools are showing up on work devices.

What is not known yet

The report notes a downside: if automated safety systems throttle permissions too aggressively, they could slow work down. The material does not say what the tools cost, when each one becomes available, or which customers can use them first.