Microsoft Unveils AI Agent Audit Standards Initiative With No Spec, Log Format or Deadline
11 hour ago / Read about 28 minute
Source:TechTimes

The Microsoft store on Fifth Avenue in Midtown Manhattan is shown June 4, 2018 in New York City. Microsoft officially announced today an agreement to buy GitHub, a code repository company popular with software developers, for $7.5 billion in stock. Getty images/Drew Angerer

Microsoft announced the AI Transparency Initiative on Tuesday at its Ignite conference, a plan with more than 40 organizations to write open standards that let companies see, independently verify and restrict what AI agents do. CEO Satya Nadella made the announcement, and the release names Accenture, Adobe, KPMG, Mastercard and Volvo Cars as participants. What it does not contain is a specification, a log format, a deadline or a ship date.

That gap matters to any security or compliance team deciding how much to trust agent audit evidence today. Microsoft says organizations should check AI activity independently with their own security and compliance systems rather than rely on the model provider's account. Yet the work is meant to flow into Microsoft's own products, and the release does not say who will operate the "independent" systems that do the checking. Buyers can use the open question as a checklist: ask where each piece of evidence comes from, who runs the system that produced it, and whether another vendor's tools can read it.

Microsoft Names Five of 40-Plus Partners but No Spec, Log Format or Deadline

The release calls for open, industrywide standards for transparency, security and governance across models and providers. Microsoft says it will build the work into Microsoft AI, Foundry and Copilot, and that organizations should be able to keep evidence when models or providers change.

ItemDetail
AnnouncedTuesday, Nov. 17, 2026, at Microsoft Ignite
Led byMicrosoft; announced by CEO Satya Nadella
Participants"More than 40" organizations; five named
Government linkExpanded Microsoft participation in a NIST agent-standards program
ProductsMicrosoft AI, Foundry and Copilot, no dates
Named standard, schema or timelineNone
Sign-upaka.ms/learnmoreAItransparency

Coverage so far traces back to Microsoft's release; Investing.com's report relies on the release and quotes no outside analyst. The partner quotes in the release endorse the goal rather than evaluate an approach. Adobe chief security officer Aanchal Gupta says security teams need visibility into what agents access and do. KPMG's John Israel, the firm's global chief information security officer, backs open, industrywide standards for responsible AI. Accenture's Harpreet Sidhu says an open, industry-driven approach gives organizations a foundation for trust. Mastercard and Volvo Cars are named without quotes.

Nadella Asked for Tamper-Proof Agent Records Five Weeks Before Ignite

The initiative follows a run of public positions from Nadella. On Sept. 13, Nadella welcomed embedded evaluators, an idea raised in Anthropic CEO Dario Amodei's essay, while saying such efforts cannot be controlled by a handful of entities. On Oct. 10, he went further in a long post on X, arguing that companies should treat powerful models like privileged insiders that may be compromised. Coverage of the post describes a framework that also includes tamper-proof, human-readable logs of AI actions and independent audits of AI deployments.

The Ignite release turns those themes into a three-part commitment of See, Verify and Control, but it stops at principles. Nadella's two public tests, records no one can quietly alter and oversight no small group controls, are the ones the initiative will eventually have to meet.

How Does Microsoft Define Independent Evidence for AI Agents?

Microsoft's framing starts from a visibility problem: organizations can observe much of the activity around AI models but often cannot reconstruct what happened between a request and its outcome. Under the Verify goal, evidence should identify whether it came from the model, runtime or independent systems such as identity and policy, and model-reported signals should be checked against independently observed records.

That three-tier provenance idea is the technical core. An agent can narrate its own steps, but the narration is only as reliable as the model producing it. Records written by systems outside the model, such as the identity provider that issued the agent's credentials or the policy engine that approved a tool call, can serve as a cross-check. Nadella's call for tamper-proof records points the same way: entries that cannot be edited or deleted after they are written.

The constraint is completeness. A log that is tamper-proof but silently missing events still fails an audit, which is why the next case matters.

Copilot Audit Logs Once Missed File Summaries, Researcher Says Microsoft Stayed Quiet

Pistachio CTO Zack Korman reported that on July 4, 2025, he found he could keep Microsoft 365 Copilot from logging file-summary interactions simply by asking for a summary without a link. Microsoft fixed the issue shortly before The Register's report in August 2025 and rated it "important," and The Register said no CVE was issued. The outlet reported that Microsoft chose not to notify customers, and Korman said organizations that used Copilot before Aug. 18 may have incomplete audit logs. A Microsoft spokesperson said the company appreciated the researcher's findings, which the outlet said did not directly answer its questions.

Korman also noted that Zenity chief technology officer Michael Bargury had raised it at Black Hat in August 2024, a year before the fix. The case cuts against the principle the initiative states: Microsoft's own logging product once missed an agent action, and the researcher's account says customers were not told. It is a single historical incident, not evidence about today's Agent 365, but it shows why independent verification needs a spec that says what must be logged.

Who Runs the Identity and Policy Systems That Count as Independent?

The release does not say. Microsoft sells Entra, Purview, Defender and Agent 365, the products that supply identity, policy and monitoring records, so the "independent systems" in Microsoft's own framework may be Microsoft products. Agent 365 went generally available on May 1, 2026, at $15 per user per month or bundled in the Microsoft 365 E7 suite at $99. If the independent tier turns out to be Microsoft's own, verification would be a licensed feature of the vendor being verified.

Independent analysts raised the same structural concern about Microsoft's Sept. 14 Code of Conduct for its MAI models. In Computerworld's September 14 report, Thomas Randall, research director at Info-Tech Research Group, said Microsoft defines the thresholds, selects the evaluators and gives its own executives the final deployment call, and he asked for evaluators with continuous access and audit logs. Frank Dickson of Dickson Research called the document "shy on mechanism," noting it names no auditor, no verification method and no stated consequence for a violation. Those critiques concern the Code, not the initiative, but they describe the test the initiative has yet to pass.

NIST Agent Standards Initiative Has Three Pillars and Now Carries an SI Label

Microsoft says it will expand its participation in a National Institute of Standards and Technology agent-standards program and use that research to shape open standards. The release calls it the "SI Agent Standards Initiative," and NIST's program page now carries the same label, expanding it as Super Intelligence, with a last-updated date of Nov. 9, 2026. See the SI Agent Standards Initiative page for the current wording. NIST's February announcement used the name AI Agent Standards Initiative.

The program has three pillars: facilitating industry-led standards, fostering community-led protocols and investing in research. NIST also ran a request for information on agent security that closed March 9, and an identity and authorization concept paper was open for comment through April 2. NIST's page describes its output as voluntary guidelines, not law. The release does not say which NIST workstreams Microsoft will join or how its work maps to them.

Agent 365 Already Governs Agents at KPMG, and the Initiative Does Not Say How They Connect

Microsoft introduced Agent 365 at Ignite on Nov. 18, 2025, as a control plane for deploying and governing agents. The new release does not say how the initiative relates to it. KPMG, a named participant, said on June 9 that it would begin adopting Agent 365 globally and extend Microsoft 365 Copilot to its more than 276,000 professionals. A KPMG executive is also one of the quoted partners, which means a participant is both a large customer and an endorser.

Whether evidence from Agent 365 today is agent-emitted, runtime-emitted or independently generated is the practical question the initiative leaves open. Microsoft has not published a mapping, and no log format or export schema accompanies the announcement.

Buyers Can Ask Four Questions of Any Vendor Before Standards Exist

Until a specification is published, the open items in the release translate into four procurement questions. Which component produced each audit record: the model, the runtime, or an identity or policy system? Who operates and audits the systems that count as independent? Are records tamper-evident and complete, and how does the vendor prove completeness? Can the evidence be exported to another vendor's security tools without losing fields?

Watch for the full participant list, a governance structure, the first draft specification and where it is hosted, how the work lines up with NIST's identity and authorization efforts, and whether other model and cloud vendors join.

Frequently Asked Questions

What is the Microsoft AI Transparency Initiative?

It is a Microsoft-led effort announced Nov. 17, 2026, at Ignite with more than 40 organizations to develop open standards for seeing, verifying and controlling AI agent activity across models and providers. It has no published specification or timeline yet, according to the AI Transparency Initiative release.

Are the checks in the initiative really independent of Microsoft?

That is not yet answerable. The framework says evidence should come from systems outside the model, such as identity and policy, but the release does not say who runs those systems. Today's governance products that supply such records are sold by Microsoft.

Which organizations are in the initiative?

Microsoft names Accenture, Adobe, KPMG, Mastercard and Volvo Cars. The other participants in the "more than 40" total have not been named in the release.

How does NIST's agent standards program fit in?

Microsoft says it will expand its participation in NIST's agent-standards program and use the research to shape open standards. The SI Agent Standards Initiative has three pillars covering industry-led standards, community-led protocols and research.