Microsoft Defender adds new controls to keep employees’ AI agents in check

Microsoft has introduced new security controls for AI agents that have access to an organisation’s sensitive data and critical systems. Companies can now see which AI agents their employees are using, monitor what those agents do, and block risky decisions.

Why it matters

Microsoft security chief Hayete Gallot said the most advanced AI models cannot be fully trusted. They are good at completing specific tasks, but not always reliable at applying judgement that fits an organisation’s rules and circumstances, even though they are given access to important data and digital tools.

What the new features do

• Detect 60 kinds of AI agents, including Claude Code, Codex, and OpenClaw
• Limit an agent’s permissions to data and systems
• Require human approval before risky decisions

The features have been added to Microsoft Defender, the company’s security service.

There is a business angle too

Microsoft has its own interests here. It is an investor in OpenAI and Anthropic and builds AI models itself. Selling AI-agent security products can also draw organisations further into its services.

Across the industry, efforts to control AI-agent risks are growing. More than 100 technology companies, Microsoft among them, have backed Nvidia’s new open agent safety platform, which monitors agents in a secure environment and can stop them when needed.