"What the model tells us about itself is not independent proof." Satya Nadella said this on 17 November 2026 in San Francisco, opening Ignite, Microsoft's conference for businesses, according to the edited version of his speech published on the group's official blog. The CEO then announced the AI Transparency Initiative. The press release mentions more than 40 associated organisations and names five: Accenture, Adobe, KPMG, Mastercard and Volvo Cars. ActuIA reported it on 18 November in a brief.
The announcement closes an autumn during which Microsoft sold IT departments the idea that an AI agent monitors itself like a high-risk employee. The controls delivered so far observe what surrounds the model, and the company itself writes that they are not enough.
The initiative has three components and no dated deliverables
The first component goes through Washington. Microsoft says it wants to expand its participation in the NIST initiative, the US National Institute of Standards and Technology, on agent standards. This programme was launched on 17 February 2026 by CAISI, the NIST centre responsible for AI standards. It aims for voluntary guidelines and industry-led protocols. As of 19 November, its page, updated on 9 November, only shows three participation methods whose deadlines, 9 March, 20 March and 2 April, have passed.
The second component brings partners together around open standards for responsible AI systems that are independently verifiable and safe by design, regardless of model or provider. The third involves in-house products. Microsoft will integrate this work into Microsoft AI, Foundry and Copilot, so that customers can use the evidence produced with their security and compliance tools, from Microsoft or other providers.
The press release sets no timeline and publishes neither specification drafts nor log formats. Interested organisations are directed to a registration form. Neither OpenAI nor Anthropic, whose models Microsoft distributes, are among the five names cited.
Microsoft proposes to confront the model's narrative with other records
The press release is organised around three verbs: see, verify and control. "Agents cannot be the sole source of truth about their own behaviour," writes Microsoft, which also argues that one cannot rely on frontier models to govern themselves. Each piece of evidence about the agent's identity, the intent delegated to it and its actions should indicate where it comes from: the model, the execution environment around it, or independent systems such as those managing identity and rules. What the model declares could then be compared with records observed elsewhere. The company should be able to carry out this verification with its own tools, without relying solely on the model provider's account of its behaviour.
The mechanism itself is described in the conditional. The text does not say who operates the independent systems, nor whether a third party audits the whole. On 10 October, in a personal text, Satya Nadella asked that every significant action of a model leave proof that is "tamper-proof and human-readable". The word does not appear in the 17 November press release.
In Microsoft's offering, agent identity and conditional access fall under Entra, a company product. In Microsoft 365, independent records would therefore come from the same provider as Copilot. The independence described is that of the control system vis-à-vis the model. It extends to the publisher only if the evidence format is open and readable by competing tools, which the press release promises without having published the format.
Only part of the October controls reaches general availability
On 19 October, the company detailed its doctrine: an identity and a register for each agent, a continuously recalculated risk level, and rights that tighten. ActuIA noted then that the post gave neither status nor date for these new features, and that the appointment was set for Ignite.
On 17 November, what reaches general availability involves adjacent or older building blocks. Project Perception, opened in October by invitation to a small number of customers, becomes Microsoft Perception and reaches general availability, according to the conference's Book of News. ActuIA reported it in a brief. Runtime protection for agents installed on Windows PCs, in preview a month earlier, is generally available for a list of named agents, including Claude Code and GitHub Copilot CLI. It is intended for Defender for Endpoint P2 customers. Automated lifecycle management of agent identities is also generally available: it verifies that each agent has a human sponsor, and a default rule disables identities inactive for 60 days, details the Entra blog.
The core of the October system is only in preview. Dynamic Agent Risk, the risk score, is in preview in Agent 365, according to the Book of News. The new Perception agents and playbooks tasked with confining AI agents remain in limited preview, writes the Microsoft Security blog. The same applies to dynamic permissions in Entra ID Governance, which grant limited access to a task and a duration, and dynamic conditional access controls, which adjust protections when risk changes. Microsoft does not publish the risk thresholds that trigger a restriction. The deployment of Copilot Autopilot and its agent identity, still in preview, was detailed by ActuIA on 18 November.
The same security post sets the limit of these tools. They allow securing and confining agents, but they "will not give [customers] visibility into how the model behind an agent behaves". Without this visibility, Microsoft says it is difficult to say, after a consequential action, what was requested and authorised from the AI, what happened between the request and the result, and whether it can be independently verified. A month earlier, the 19 October post announced that signals from the model, including its intermediate reasoning, would feed Dynamic Agent Risk. Satya Nadella wrote as early as 10 October that the transparency of a model's reasoning is, on its own, neither sufficient nor reliable.
The press release leaves responsibility with the customer
Responsibility remains with the organisation, writes the press release, adding that blaming AI is not an answer and trusting the model is not a control. The text mentions neither the European Union nor the AI Act. For the record, the European regulation requires, in paragraph 6 of its Article 26, deployers of high-risk AI systems to keep automatically generated logs, insofar as they are under their control, for a period appropriate to the system's purpose and at least six months, unless otherwise provided by Union or national law. The obligation only applies to uses classified as high-risk, and Microsoft claims no compliance.
At this stage, the initiative is a finding and a method, with no associated product. Its three sources of evidence already make it possible to question any agent provider, including Microsoft, about the component that produced each audit record, the one that uses it, and the possibility of exporting these records without loss to a third-party security tool.
As of 19 November, the eve of Ignite's closing, ActuIA had found from Microsoft, on the initiative, only the press release, the speech and its republications, a passage from the security blog and a registration form.
Quotes translated from English by ActuIA.
Our articles will then appear first in Google Top Stories.
