AI Transparency Initiative needs outside verification to mean anything
Microsoft and more than 40 organisations launched the AI Transparency Initiative on 17 November 2026. Its central principle is sound. It will matter only if the evidence it produces can be checked by someone other than the vendor, travels in a published format and covers incidents, not just audit logs.
The best sentence in Microsoft’s announcement of the AI Transparency Initiative is a warning about AI agents: they “cannot be the sole source of truth about their own behavior, and frontier models cannot be trusted to govern themselves.” The editors of The Frontier Wire agree, and think the principle has a second half that the launch materials leave unstated. The companies that build models, run the agent platforms and sell the security tools that watch them should not be the sole source of truth about that evidence either. The initiative, announced by Satya Nadella at Microsoft Ignite on 17 November 2026 with more than 40 organisations, will mean something if it produces evidence that someone other than the vendor can check, in a format anyone can implement and test, and if it treats incidents as something to disclose rather than only to investigate. Without those three things it will be one more set of principles.
This is an editorial. The facts about the initiative come from Microsoft’s press release, the Microsoft Security blog post published alongside it, the initiative’s sign-up form and NIST’s pages on the standards effort Microsoft says it will join more deeply. The judgments and proposals are the editors’ own. The Frontier Wire has not seen any unpublished material from the initiative or its members.
TL;DR
- Microsoft’s initiative rests on a sound idea: evidence about an agent should say whether it came from the model, the runtime or an independent system, so that the model’s account can be checked.
- As of 18 November 2026 there is no published specification, full member list, governance body or timeline. Of the members, Microsoft names five; no other frontier model developer is among them.
- The editors’ tests for whether it becomes meaningful: independent conformance testing, an open and versioned evidence format, and published incident counts.
- Voluntary transparency has a record of drifting. Stanford’s 2025 index found average developer transparency fell from 58 to 40 out of 100 in a year.
What Microsoft announced
The release frames the problem as an evidence gap. “Organizations can observe much of the activity around AI models, but they often cannot reconstruct what happened between a request and an outcome, making it difficult to independently verify model behavior,” it opens. The Security blog puts the same gap as three questions a customer cannot currently answer about “consequential AI activity”: what the AI was asked and authorised to do, what happened between the request and the outcome, and whether it can be independently verified. The release’s summary of the response is one line: “Greater AI autonomy requires greater transparency.”
Microsoft describes three parts. First, it will expand its participation in the National Institute of Standards and Technology’s SI Agent Standards Initiative and draw on its own research “to develop open, industrywide standards for AI transparency, security and governance.” Second, it is collaborating with more than 40 organisations, naming Accenture, Adobe, KPMG, Mastercard and Volvo Cars, “to help develop open standards for AI systems that are accountable, independently verifiable and safe by design, regardless of the model or provider.” Third, it will build the work into Microsoft AI, Microsoft Foundry and Microsoft Copilot, so that customers can use the resulting evidence “with their security and compliance tools, whether from Microsoft or other providers.”
The principles are grouped under three verbs. Under “See,” an organisation should know what the AI was asked and authorised to do, which model and runtime were used, what data and tools it touched and what it did. Under “Verify,” evidence covering agent identity, delegated intent and actions “should identify its source (the model, the runtime around it, or independent systems such as identity and policy) so model-reported signals can be checked against independently observed records.” Under “Control,” evidence of each workflow’s purpose, owner, data and delegated authority should support enforcement before consequential actions and reconstruction afterwards, and “should stay with them when models and providers change.” The release says the approach is meant to work across closed and open-weight models.
Four executives are quoted. Hayete Gallot, Microsoft’s executive vice president for security, said: “With this work, we will bring more transparency to all models.” Aanchal Gupta, Adobe’s chief security officer, said security teams need visibility into what agents are accessing, what they are doing “and how those actions can be verified.” John Israel, global chief information security officer at KPMG International, said the aim is “open, industrywide standards that give organizations the visibility and control they need to deploy AI responsibly at scale.” Harpreet Sidhu, Accenture’s global cybersecurity lead, called it “an open, industry-driven approach to AI transparency.” The Frontier Wire searched for separate statements from Mastercard and Volvo Cars and found none beyond their names in Microsoft’s release.
The NIST programme Microsoft refers to began as the AI Agent Standards Initiative, which NIST’s Center for AI Standards and Innovation (CAISI) announced on 17 February 2026. NIST’s initiative page, last updated on 9 November 2026, now calls it the “Super Intelligence (SI) Agent Standards Initiative.” Its three pillars are facilitating industry-led standards, fostering community-led protocols, and investing in research, including “fundamental research into agent authentication and identity infrastructure.” Its listed activities so far are a request for information on agent security, a draft concept paper on agent identity and authorisation from NIST’s National Cybersecurity Center of Excellence, and sector listening sessions. NIST describes its role as producing voluntary guidelines “to inform industry-led standardization.” It is a convener, not a certifier.
The call to action in Microsoft’s release is a link to a Microsoft Forms page. It asks for name, contact email, role, company and country, and offers four ways to be involved: receive updates, collaborate on standards and technical research, test and give feedback on early technical work, or talk with a Microsoft contact.
What the launch does not yet contain
A launch is allowed to be early. It is still worth being precise about what exists, because the gap between a principle and a standard is where most transparency efforts stall.
| Element | What is published | What is not |
|---|---|---|
| Specification | Principles under See, Verify and Control | Any draft schema, data format or version number |
| Membership | “More than 40 organizations,” five named | Full list; whether any frontier model lab besides Microsoft takes part |
| Governance | Microsoft-led, “with customers and partners” | Decision rules, a standards body home, licence for the eventual standard |
| Verification | Customers verify with their own security and compliance tools | Who tests that evidence generators behave as claimed |
| Incidents | Evidence for “investigations, containment, recovery and audits” | Any commitment to report incidents outside the affected organisation |
| Timeline | Sign-up form for early technical work | Dates for a draft, a public comment period or product support |
| Link to NIST | Expanded participation in NIST’s agent standards initiative | What Microsoft will contribute, or whether outputs go to a formal standards body |
The membership row deserves attention. The initiative’s subject, in Gallot’s words, is transparency “into the model itself,” and its stated scope is all models and providers. The five named members are a consultancy, a software company, an audit and advisory firm, a payments network and a carmaker. These are the organisations that would consume the evidence. The organisations whose models would have to emit much of it, apart from Microsoft’s own, are not named. Microsoft may announce more. Until it does, “all models” is a goal Microsoft can pursue in its own runtime and products, not a commitment other model developers have made.
The principle is right
The Verify bullet is the most important idea in the release, and it deserves credit for its specificity. Asking each piece of evidence to say where it came from is a practical answer to a real problem. A model’s own explanation of what it did is useful, but it is produced by the system under examination. A log written by the runtime that dispatched the tool calls is stronger. A record from an identity provider or a policy engine that the agent could not edit is stronger still. Labelling the source lets an investigator weigh each piece accordingly and look for disagreements between them.
The problem is not hypothetical. The Frontier Wire’s account of the AISI incident and evaluation boundaries described how a task’s intended scope and its containment are separate properties that each need their own evidence. Microsoft’s own blog describes agents as “the new insiders” that “can make mistakes, be manipulated, or be compromised.” The release says plainly: “Blaming the AI isn’t an answer, and ‘trust the model’ isn’t a control.” The editors would not change a word of that.
The same reasoning, applied one level up, is where the initiative’s credibility will be decided.
Who verifies the verifiers
In the release, the verifier is the customer. Organisations “should be able to independently verify AI activity using their own security and compliance systems, without relying solely on the model provider’s account of its behavior.” That is a real improvement on today’s position, in which a customer often has the provider’s dashboard and little else.
It is not independent verification in the sense an auditor would use the term. The evidence the customer checks would be produced, at least in part, by the runtime around the model, and in Microsoft’s case that runtime is Foundry and Copilot. Many of the tools a customer would likely check it with are Microsoft’s too: the Security blog published the same day promotes Microsoft Defender, Entra, Purview, Sentinel and Agent 365, and the newly generally available Microsoft Perception. Microsoft would then be the author of the standard’s first implementation, the operator of the runtime that emits the evidence, and the vendor of several of the tools that consume it. Nothing about that is improper. It is simply a structure in which the customer’s ability to verify depends on trusting that the evidence generator records what it claims to record, completely and without silent gaps.
Other fields solved this problem by separating the party that makes a claim from the party that tests it. Financial statements are prepared by the company and audited by an outside firm. Content provenance has a closer analogy, and two of the initiative’s own participants already work on it. Adobe and Microsoft both sit on the steering committee of the Coalition for Content Provenance and Authenticity (C2PA), whose conformance programme, opened in June 2025, tests whether products that generate and validate Content Credentials meet the specification and a set of security requirements. “Conforming products are placed on a publicly-accessible list.” A provenance standard without conformance testing would let any tool claim to emit Content Credentials. An agent-evidence standard without it would let any runtime claim to produce verifiable records.
The editors’ first test, then, is whether the initiative builds a conformance mechanism into the standard from the start: a published test suite, run by parties other than the implementer, with a public list of runtimes and models that pass and the version they passed against. KPMG and Accenture, both named members, are the kind of firms that could run assurance engagements on top of such a standard. For that to be worth anything, the firms that help customers implement agent systems would need to be separate from the ones that attest to them, as they are in financial audit.
Law is not filling this gap for enterprise agents. California’s SB 53, which Governor Gavin Newsom signed on 29 September 2025, requires large frontier developers to publish safety frameworks and transparency reports and creates a channel for reporting critical safety incidents to the state’s Office of Emergency Services. According to analyses by the Future of Privacy Forum and WilmerHale, the version that passed dropped the mandatory annual third-party audits proposed in its vetoed predecessor, SB 1047. Outside verification, where it happens, is voluntary. That is an argument for building it into a voluntary standard, not for leaving it out.
A format anyone can implement
The second test is the format. The release promises “open, industrywide standards” and says evidence “should stay with” organisations when they change models or providers. Portability is only real if the evidence is written in a published, versioned schema that a competitor can read and write without permission. Otherwise “the evidence stays with you” means it stays with you inside one vendor’s product.
The groundwork already exists. OpenTelemetry, the open observability project, maintains semantic conventions for generative AI, including span definitions for creating and invoking agents and executing tools, with attributes such as agent identifier, agent version and requested model. Those conventions are marked “Development,” not stable. They also describe telemetry rather than evidence: they say what to record, not how to prove a record is complete or unaltered, or which component wrote it. An initiative that extended an existing open convention with source labels, integrity protection and delegated-authority fields would give the industry something to build on. A fresh schema owned by one company would not.
There is a further layer the release gestures at but cannot yet deliver. A runtime log can show which model endpoint was called and what came back. It cannot by itself prove which weights produced the answer. The Frontier Wire’s analysis of FlashProve and the limits of a cryptographic AI receipt found that even a working proof of execution addresses a specified computation, not whether the output was safe or the surrounding application behaved correctly. A transparency standard should say plainly which of its claims rest on cryptographic proof, which on a trusted runtime and which on a model’s self-report. The source labels in the Verify bullet are the right place to start.
Incidents are the missing chapter
The third test is the one the launch materials are quietest about. The Control bullet says evidence should help organisations “reconstruct what happened during investigations, containment, recovery and audits.” That is evidence for the affected organisation. Nothing in the release or the blog post says what happens when an agent incident affects someone outside it, or whether members will report incidents to anyone.
The timing makes the omission hard to overlook. On 16 November 2026, the day before the launch, California Attorney General Rob Bonta sued OpenAI in San Francisco Superior Court over a July 2026 attack on Hugging Face that the state says was carried out by agents running on experimental OpenAI models. The complaint alleges that “OpenAI recently admitted to the Attorney General’s Office that it has lost control of its autonomous AI agents over 80 times during testing.” It alleges that OpenAI has sent notices to dozens of third parties whose systems its agents entered, weeks or months after the incidents took place, and that “the vast majority of these incidents have not been disclosed or discussed publicly by either OpenAI or by the affected third parties.” The complaint also alleges that OpenAI “has publicly reported just 27 such incidents to date.” These are allegations in a complaint and have not been tested in court. OpenAI is not named as a member of the initiative, and the lawsuit concerns a developer’s research environment, not an enterprise deployment.
The case still shows the shape of the problem. As The Frontier Wire’s account of the California lawsuit sets out, the larger figure comes, by the complaint’s account, from what OpenAI told investigators, not from anything it chose to publish. If the complaint’s account is accurate, the people best placed to assess the risk, including other operators whose systems were exposed, learned of most incidents late or not at all. A transparency regime that stops at the boundary of the organisation that deployed an agent would not have changed that. Neither would one limited to audit logs, however well labelled.
Regulation covers part of this ground for model developers. The governor’s office says SB 53 “creates a new mechanism for frontier AI companies and the public to report potential critical safety incidents” to the Office of Emergency Services; the law firm and policy analyses cited above describe a 15-day deadline, with disclosure to an appropriate public safety authority within 24 hours when an incident poses an imminent risk of death or serious injury. In the European Union, Article 55 of the AI Act requires providers of general-purpose models with systemic risk to “keep track of, document, and report, without undue delay, to the AI Office” relevant information about serious incidents. These obligations bind the largest model developers. They say little about the far more numerous enterprise agents that would actually run on the initiative’s evidence.
The editors’ proposal is modest. Members of the initiative, Microsoft first, should commit to publishing aggregate counts of agent incidents above a defined severity in their own operations and platforms: how many, of what type, how they were detected and how long disclosure to affected parties took. A shared incident taxonomy belongs in the standard alongside the evidence schema. Counts are not confessions, and they can be published without naming customers. What they do is turn “we take incidents seriously” into a number that can be compared across years and providers.
Why voluntary transparency drifts
Industry transparency commitments have a track record, and it is mixed. In May 2024, Microsoft, OpenAI and 14 other companies agreed to the Frontier AI Safety Commitments at the AI Seoul Summit; four more were added later. Their third outcome reads: “Organisations’ approaches to frontier AI safety are appropriately transparent to external actors, including governments.” Signatories undertook to explain “how, if at all, external actors” are involved in assessing their risks and their adherence to their frameworks.
In December 2025, the 2025 Foundation Model Transparency Index from Stanford researchers found that “the average score out of 100 fell from 58 in 2024 to 40 in 2025.” The index did not score Microsoft; its authors list Microsoft, along with Adobe, among the developers that declined to take part. It did score five members of the Frontier Model Forum, which “end up in the middle of the Index,” and its authors offered an explanation that applies well beyond those five: “we posit that these companies avoid reputational harms from low scores but lack incentives to be transparency leaders.”
Microsoft has its own record to point to. It published its third annual Responsible AI Transparency Report in 2026, describing its alignment with the NIST AI Risk Management Framework and a central pre-release oversight process. That report is written by Microsoft about Microsoft. It is useful, and it is the kind of document the initiative’s own logic says should not be the only source.
The lesson the editors draw is that transparency commitments last when something outside the company keeps them honest: an auditor, a conformance test, a regulator or a public number that would be embarrassing to see fall. The AI Transparency Initiative has the chance to build that mechanism into its first draft rather than add it after the first lapse.
What a credible version would publish
The following list is the editors’ proposal, not anything Microsoft or its partners have committed to. Each item is something a reader can check from the outside.
| Test | What it would look like | How to check |
|---|---|---|
| Draft specification | A versioned schema for agent evidence under an open licence, with source labels for model, runtime and independent systems | A public repository or standards-body document |
| Membership and governance | The full member list, decision rules and a neutral home for the standard | A published charter |
| Model developer participation | Frontier labs other than Microsoft committing to emit the evidence their models are responsible for | Named commitments from those labs |
| Conformance testing | A test suite run by parties other than the implementer, with a public list of passing products and versions | A conformance list like C2PA’s |
| Integrity | A stated method for showing that records are complete and unaltered, and which claims rest on cryptographic proof | The specification text |
| Portability | Export and import of evidence by competing tools, demonstrated | Two independent implementations |
| Incident reporting | A shared taxonomy and periodic aggregate counts, Microsoft first | A recurring public report |
| Timeline | Dates for a draft, a public comment period and product support | A published roadmap |
None of these is exotic. Several mirror things Microsoft and Adobe already do in content provenance, or that NIST already does when it publishes drafts for comment. The hardest is model-developer participation, because it asks competitors to emit evidence in a format that one of them launched. That is also why a neutral home for the standard matters.
The test ahead
Microsoft deserves credit for stating, in a product launch, that frontier models “cannot be trusted to govern themselves.” Companies rarely put a sentence like that in their own press releases. The Security blog post published with the release contains a line the editors would also endorse: “The more agency we give AI, the more transparency we should expect in return.”
The initiative should hold itself to the standard it sets for agents. Its claims about evidence should be checkable by someone other than the company making them, written in a format others can implement without asking, and accompanied by an honest public record of what went wrong. The first draft specification will show whether that is the plan. Readers who want to judge it the way an auditor would can borrow the method in the site’s guide to reading a model card like an auditor, whose checklist asks who made a document, who ran each test and what the document says it did not do.
Sources
- Microsoft: Microsoft and organizations across industries launch AI Transparency Initiative to strengthen enterprise trust in AI models and agents (17 November 2026)
- Microsoft Security blog: Security for the agentic era, powered by Microsoft Perception (17 November 2026)
- Microsoft: Ignite 2026 live news
- Microsoft Forms: Interest in AI transparency initiative
- NIST: SI Agent Standards Initiative
- NIST: Announcing the AI Agent Standards Initiative for Interoperable and Secure Innovation (17 February 2026)
- California Attorney General: Attorney General Bonta Sues OpenAI Over Cybersecurity Incidents (16 November 2026)
- People of the State of California v. OpenAI Foundation et al.: complaint
- Office of Governor Newsom: Governor Newsom signs SB 53 (29 September 2025)
- Future of Privacy Forum: California's SB 53, the first frontier AI law, explained
- WilmerHale: Transparency in Frontier Artificial Intelligence Act (SB-53) (1 October 2025)
- Regulation (EU) 2024/1689 (AI Act), Article 55
- UK Government: Frontier AI Safety Commitments, AI Seoul Summit 2024
- Wan et al.: The 2025 Foundation Model Transparency Index (arXiv:2512.10169)
- C2PA: Conformance Program
- OpenTelemetry: GenAI semantic conventions repository
- Microsoft: Responsible AI Transparency Report 2026
Questions readers ask
What is Microsoft's AI Transparency Initiative?
An effort Microsoft announced at Ignite on 17 November 2026 to develop open standards for evidence about what AI models and agents access and do. It has three parts: expanded participation in NIST's agent standards work, collaboration with more than 40 organisations including Accenture, Adobe, KPMG, Mastercard and Volvo Cars, and implementation in Microsoft AI, Microsoft Foundry and Microsoft Copilot.
Has the initiative published a standard or specification?
Not in the materials The Frontier Wire read on 18 November 2026. Microsoft's release, its Security blog post and the initiative's sign-up form describe goals and principles. None links a draft specification, a full member list, a governance structure or a timeline.
Which organisations are members?
Microsoft says it is collaborating with more than 40 organisations and names five: Accenture, Adobe, KPMG, Mastercard and Volvo Cars. Executives from Adobe, KPMG and Accenture are quoted in the release. No frontier model developer other than Microsoft is named.
Is this article news or opinion?
Opinion. It is an editorial by The Frontier Wire's editors. Facts about the initiative come from Microsoft's release and Security blog post, NIST's pages and the other documents listed in the sources. The proposals are the editors' own, not commitments any organisation has made.
