Microsoft wants AI agents audited by records the model doesn’t write
Microsoft and more than 40 organizations will draft open standards so companies can check what AI agents did using records the model does not control.

Microsoft used its Ignite conference on November 17 to launch the AI Transparency Initiative, working with more than 40 organizations on open standards that let companies independently verify what AI agents did.
CEO Satya Nadella announced it, according to Microsoft. The company’s argument is that firms can see much of the activity around a model but often cannot reconstruct what happened between a request and an outcome.

Three parts, and no dates for any of them
Microsoft describes the initiative as having three strands. First, it will widen its participation in the National Institute of Standards and Technology’s AI agent standards effort and use its own research to help write open, industrywide standards for AI transparency, security and governance. (The release prints the name as “SI Agent Standards Initiative”; NIST’s program is called the AI Agent Standards Initiative, so this looks like a typo.)
Second, Microsoft is working with more than 40 organizations, among them Accenture, Adobe, KPMG, Mastercard and Volvo Cars. The goal is standards for systems that are accountable, independently verifiable and safe by design, whatever the model or provider.
Third, Microsoft says it will build the results into Microsoft AI, Microsoft Foundry and Microsoft Copilot. Customers could then feed the resulting evidence into security and compliance tools from Microsoft or other vendors.
The announcement gives no release dates, no product names for the features and no pricing. Organizations that want to take part or follow progress can register at aka.ms/learnmoreAItransparency. For now, nothing is available to deploy.
See, verify, control, with the model as a witness
Microsoft frames the work as three capabilities. See means knowing what an agent was asked and authorized to do, which model and runtime it used, and which data and tools it touched. Control means keeping the purpose, owner, data and delegated authority of each workflow on record, so policies apply before consequential actions and investigators can reconstruct events afterward.
Verify is the distinctive part. Microsoft’s position is that an agent should never be the single authority on what it did. Each piece of evidence about identity, delegated intent and actions should say where it originated: the model itself, the software environment wrapped around it, or separate systems that handle identity and policy. That way, whatever a model reports about itself can be checked against records it did not produce. The company says the approach should cover closed and open-weight models, so teams can swap the model behind an agent without breaking security governance or audit.
Hayete Gallot, Microsoft’s security chief, said “for security controls to be effective, every organization needs greater transparency into the model itself.” The release also states that “trust the model” is not a control.
A standards bet that leans on an unfinished federal effort
The angle here is rules and safety. The initiative ties itself to NIST’s AI Agent Standards Initiative, which NIST’s Center for AI Standards and Innovation announced on February 17, 2026. According to NIST, the center leads it with the agency’s Information Technology Laboratory and the National Science Foundation. At launch NIST opened a request for information on agent security, due March 9, and a concept paper on agent identity and authorization, due April 2. Both deadlines have passed. In the sources reviewed, one third-party guide, undated, said NIST had not yet published a finalized standard under the program, and no newer official deliverables turned up. Microsoft’s standards will therefore be written alongside a federal process that is still gathering input.
The vendor angle also deserves a note. Microsoft sells an agent platform and a security portfolio, so it benefits if the evidence format that emerges plugs into its own products. The release says the output should work with security tools from other providers, but that is a stated intent, and the source gives no detail on governance, who will own the specification, or whether rival model makers are among the 40-plus participants. The quotes in the release come from Microsoft, Adobe, KPMG and Accenture; none of the claims has been independently verified.
Administrators are already getting controls of a narrower kind. Our earlier report on GitHub’s policy that lets Copilot admins bar MCP servers and skills that do not come from plugins covers restrictions on what an agent may load. Microsoft’s initiative aims at the other end: proving afterward what an agent actually did.
What readers can do now
Practically, only sign up. Security and compliance teams running agents today have no new tooling from this announcement, and Microsoft names no timeline for Copilot, Foundry or Microsoft AI. The accountability point is the one concrete commitment: Microsoft says responsibility stays with the organization, and leaders answer to customers, boards, auditors and regulators when something goes wrong.
Sources
Sign Up for the Daily AI Pulse
One email a day. All the stories that matter.
By signing up, you agree to our Terms of Use and acknowledge the data practices in our Privacy Policy. You may unsubscribe at any time.
Alex CarterAI News & Big Tech CorrespondentAlex Carter writes for Aiholics, keeping readers updated on the fast-paced world of AI and Big Tech.
You may also like

Microsoft Copilot Studio: What you can build, what it costs and how to start

Microsoft folds its customer AI perks into Frontier Accelerate, with no dollar figures

Microsoft’s Copilot Cowork team merged 1,211 PRs in a day, then swapped Python for Rust
