Microsoft adds controls that start AI agents with no access

Microsoft tells of new tools for protection that see AI agents as a new type of insider risk. Nadim Abdo, a vice president of Microsoft, wrote this in a blog post. The tools extend Project Perception, the agent protection system that Microsoft started in the first part of this year. Microsoft adds 35 types of local agents, for a total of 60. It gave no test results.

How it works

Agents start with no authority. They get only the access that an approved task must have, and only while the task is in work. A risk score uses signals from identity, data, network and the work of the agent.

If risk goes up, policy can let reads continue but cause a blockage of writes. It can also tell a person to give approval, or isolate and stop the agent.

The numbers

The 60 types of agents include Claude Code, OpenAI Codex and GitHub Copilot. Companies can put them in one registry to see which person has each. Terrell Cox, a vice president of Microsoft, tells that Microsoft operates hundreds of thousands of agents. Microsoft gave no test results for the new controls.

What is not known

Microsoft will add chain-of-thought signals, the written reasoning steps of a model, to the risk score. The source gives no dates and no prices. It also does not tell which parts a person can use at this time. Microsoft tells that decisions about protection must not use the model.

Insight: Microsoft wants to own the control layer for all agents

On 8 October, Microsoft made its MXC sandbox available to all. Microsoft adds agent identity, risk scores and a registry of agents from rivals, such as Claude Code and OpenAI Codex. This reads as a bid to hold the control layer above each agent, with a different model in it. Who holds the registry can set the rules. Monitor if customers accept one vendor that makes agents and also monitors them.