At Microsoft’s Ignite conference on November 17, 2026 in San Francisco, Satya Nadella announced the AI Transparency Initiative, aimed at making AI agents—programs that act on behalf of a user—verifiable. The group claims over 40 organizations are involved. However, the press release sets no timeline and publishes no dated deliverables.
The CEO summarized the company’s position as follows: “What the model tells us about itself is not independent evidence.” The press release specifically names five partners: Accenture, Adobe, KPMG, Mastercard, and Volvo Cars.
The initiative has three parts. The first involves Washington, where Microsoft wants to expand its participation in a NIST (the US National Institute of Standards and Technology) initiative dedicated to agent standards. Launched on February 17, 2026 by CAISI, NIST’s center for AI standards, this program aims for voluntary guidelines. On its page, updated November 9, the participation deadlines—March 9, March 20, and April 2—have already passed.
The second part brings partners together around open standards for independently verifiable AI systems, regardless of model or provider. The third concerns in-house products: Microsoft AI, Foundry, and Copilot are to enable customers to leverage evidence generated with their security and compliance tools, whether from Microsoft or other providers.
Confronting the Model’s Narrative with Other Records
The press release hinges on three verbs: see, verify, and control. Microsoft writes that “agents cannot be the sole source of truth about their own behavior.” According to the text, each piece of evidence concerning the agent’s identity, the intent delegated to it, and its actions should indicate its origin: the model, the execution environment, or independent systems, such as those managing identity and rules. What the model claims could thus be compared to records established elsewhere.
The mechanism is still described conditionally. The press release does not specify who operates these independent systems or if a third party audits the whole. In Microsoft’s offering, agent identity falls under Entra, a product from the publisher, meaning the announced independence mainly concerns the control system relative to the model, not the publisher itself. On October 10, in a personal text, Satya Nadella called for every significant action by a model to leave “unforgeable and human-readable” evidence. This term does not appear in the November 17 press release.
Only Part of the Controls Are Available
Several building blocks announced in October are now generally available, meaning accessible to all customers. Project Perception becomes Microsoft Perception, according to the conference’s Book of News. Runtime protection for agents installed on Windows endpoints is available for a named list of agents, including Claude Code and GitHub Copilot CLI, and is reserved for Defender for Endpoint P2 customers. Automated agent identity management, which verifies that a human sponsor is associated with each agent, is also available.
The core of the system remains in preview. Dynamic Agent Risk, the risk score, is in Agent 365, still in preview according to the Book of News. The dynamic permissions of Entra ID Governance and the dynamic controls for conditional access are also in preview. Microsoft does not publish the risk thresholds that trigger a restriction.
The November 17 security blog post also acknowledges a limitation: these tools allow securing and containing agents, but they will not give customers visibility into the operation of the model behind an agent. The text also does not say how companies will be able to independently verify what was asked of the AI after a high-consequence action.
The press release further specifies that responsibility remains with the organization deploying the agent. It mentions neither the European Union nor the AI Act. This European regulation stipulates, in Article 26, that deployers of high-risk AI systems retain automatically generated logs for at least six months, unless otherwise provided by Union or national law.
