Microsoft’s security boss says frontier AI models can’t be trusted, no matter who built them. So Microsoft now treats AI agents like employees who could make mistakes or get hacked.
If you run Claude Code or OpenAI’s Codex on a work laptop, Microsoft’s tools can now find it. They can spot 60 kinds of agents.
Each agent gets its own ID and starts with zero permissions. It only gets what one task needs, for as long as the task runs.
- If an agent drifts off task, it can lose the power to change anything. At the highest risk, Microsoft shuts it down.
- A team of Guardian Agents digs in when an agent’s risk jumps and can lock it down on its own.
- Microsoft hasn’t said when it ships or what it costs.
It’s all built on a Satya Nadella essay that called AI models insider risks. Ted Lieu, a congressman, replied saying he largely agrees and wants his bipartisan AI kill switch bill passed this year.
