Skip to content

Microsoft recruits more than 40 groups to make AI agents auditable

News
An unmarked server enclosure behind a glass partition in a bright enterprise control room.
AI-generated editorial illustration of visibility and oversight in enterprise AI infrastructure.
Microsoft has brought more than 40 groups into a new AI effort. They will work on shared ways to check what AI agents do. The aim is to give firms records they can check for themselves.
The company launched the AI Transparency Initiative at Ignite on November 17. Partners include Accenture, Adobe, KPMG, Mastercard and Volvo Cars.

Check what AI agents did

AI agents can use files, tools and work accounts. Microsoft wants firms to track each step. Records would show what the user asked the agent to do. They would also show which model ran and which files it read.
The plan calls for records of the tools it used. Firms could then check which actions took place. Each record would show where the evidence came from.
A model may report its own actions. Software that runs the model can also keep a log. Access and policy tools can supply their own records too.

Keep records when models change

The group plans to work across models and providers. That includes closed models and models with weights users can download. Firms would keep their evidence when they switch the model.
Microsoft will also expand its work with NIST. The U.S. body runs an agent standards program. It studies ways to check an agent’s identity and access. It also supports shared protocols and security tests. NIST seeks shared rules so AI agents can work across systems on a user’s behalf.

Bring the work into products

Microsoft plans to use the work in Microsoft AI, Foundry and Copilot. Firms could feed the records into their own security tools. The work also aims to help staff trace a fault, contain it and recover. They could use tools from Microsoft or other firms.
The company has yet to set a date for a final shared standard.
Our report on Akamai's Agent Sandbox explains access rules for AI code. We also covered agents testing a live website without consent.
loading

Loading