Friday, November 20, 2026

Top 5 This Week

Related Posts

Microsoft launches initiative to make AI agents verifiable


“What the model tells us about itself is not independent evidence.” Satya Nadella uttered this sentence on November 17, 2026 in San Francisco, while opening Ignite, Microsoft’s conference for businesses, according to the edited version of his speech published on the group’s official blog. The CEO subsequently announced the AI ​​Transparency Initiative. The release talks about more than 40 associated organizations and names five: Accenture, Adobe, KPMG, Mastercard and Volvo Cars. ActuIA reported this on November 18 in a brief.

The announcement caps a fall during which Microsoft sold IT departments the idea that an AI agent monitors itself like an employee at risk. The controls delivered so far observe what surrounds the model, and the editor himself writes that they are not enough.

The initiative has three components and no dated deliverables

The first part goes through Washington. Microsoft says it wants to expand its participation in the NIST, the American Institute of Standards and Technology, initiative on agent standards. This program was launched on February 17, 2026 by CAISI, the NIST center responsible for AI standards. It aims for voluntary guidelines and protocols supported by the industry. As of November 19, its page, updated on November 9, displays only three participation methods whose deadlines, March 9, March 20 and April 2, have passed.

The second part unites partners around open standards for responsible, independently verifiable and safe by design AI systems, regardless of model or vendor. The third involves in-house products. Microsoft will integrate this work with Microsoft AI, Foundry, and Copilot, so customers can leverage evidence produced with their security and compliance tools, from Microsoft or other vendors.

The release does not set a timetable and does not publish a draft specification or journal format. Interested organizations are directed to a registration form. Neither OpenAI nor Anthropic, whose models Microsoft distributes, are among the five names cited.

Microsoft suggests comparing the model’s story to other records

The press release is organized around three verbs, see, verify and control. “Agents cannot be the sole source of truth about their own behavior,” writes Microsoft, for whom state-of-the-art models cannot be relied upon to govern themselves either. Every piece of evidence about the agent’s identity, the intent delegated to it, and its actions should indicate where it comes from: from the model, from the execution environment around it, or from independent systems like those that manage identity and rules. What the model declares could then be compared with recordings 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 behavior.

The mechanism is described in the conditional. The text does not say who operates the independent systems, nor whether a third party audits the whole. On October 10, in a personal text, Satya Nadella asked that each significant action of a model leave proof “unfalsifiable and readable by a human”. The word does not appear in the press release of November 17.

In Microsoft’s offering, agent identity and conditional access come under Entra, a product from the publisher. In Microsoft 365, independent recordings would therefore come from the same provider as Copilot. The independence described is that of the control system with respect to the model. It only extends to the publisher if the format of the proofs is open and readable by competing tools, which the press release promises without having published the format.

Only part of the October checks becomes general availability

On October 19, the publisher detailed its doctrine: an identity and a register for each agent, a level of risk continuously recalculated, rights which are tightened. ActuIA then noted that the post gave neither status nor date for these new features, and that the meeting was set at Ignite.

On November 17, what goes into general availability is due to neighboring or older bricks. Project Perception, opened in October by invitation to a small number of customers, becomes Microsoft Perception and goes to general availability, according to the conference Book of News. ActuIA reported on it in a brief. Runtime protection for agents installed on Windows workstations, 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 agent identity lifecycle management is also generally available: it verifies that each agent has a human sponsor, and a rule provided originally deactivates identities that have been inactive for 60 days, details the Entra blog.

The heart 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. Perception’s new agents and playbooks for containing AI agents remain in limited preview, writes the Microsoft Security blog. So are Entra ID Governance’s dynamic permissions, which grant task- and duration-limited access, and dynamic conditional access controls, which adjust protections as 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 November 18.

The same security ticket sets the limit of these tools. They make it possible to secure and confine the agents, but they “will not give [aux clients] visibility on how the model behind an agent behaves. Without this visibility, according to Microsoft, it is difficult to say, after a serious action, what was requested and authorized from the AI, what happened between the request and the result, and whether this can be independently verified. A month earlier, the October 19 post announced that signals from the model, including its intermediate reasoning, would power Dynamic Agent Risk. Satya Nadella wrote on October 10 that the transparency of a model’s reasoning is, in itself, neither sufficient nor reliable.

The press release leaves the responsibility to the customer

The responsibility remains that of the organization, writes the press release, which adds that blaming AI is not an answer and that trusting the model is not control. The text does not mention the European Union or the AI ​​Act. As a reminder, the European Regulation requires, in paragraph 6 of its Article 26, deployers of high-risk AI systems to keep automatically generated logs, to the extent that they are under their control, for a period appropriate to the purpose of the system and at least six months, unless otherwise provided by Union or national law. The obligation only targets uses classified as high risk, and Microsoft does not claim any compliance.

At this stage, the initiative is an observation and a method, without an associated product. Its three sources of evidence already allow any agent vendor, including Microsoft, to be questioned about which component produced each audit record, who is using it, and whether these records can be exported losslessly to a third-party security tool.

As of November 19, the day before Ignite closed, ActuIA had found at Microsoft, on the initiative, only the press release, the speech and its repetitions, a passage from the security blog and a registration form.

Quotes translated from English by ActuIA.

Our articles will be prioritized in Google’s “Featured” section.

Shredder Smith
Shredder Smith
Shredder Smith is the lead curator and digital persona behind topaitools4you.com, an AI directory dedicated to "shredding" through industry hype to identify high-utility software for everyday users. Smith positions himself as a blunt, no-nonsense reviewer who vets thousands of emerging applications to filter out overpriced "wrappers" in favor of tools that offer genuine ROI and practical productivity. The site serves as a watchdog for the AI gold rush, providing categorized rankings and transparent reviews designed to help small businesses and creators navigate the crowded tech landscape without wasting money on low-value tools.

LEAVE A REPLY

Please enter your comment!
Please enter your name here

Popular Articles