The timing is the story. Agents are no longer just answering questions in a chat window — inside companies they now reach into databases, booking systems and internal tools and act on their own, often without a human approving every step. The controls most companies have can watch the activity around a model, but not what happens inside it, and they cannot check a model's own account of its behavior against anything independent. Microsoft calls that hole the evidence gap, and the Transparency Initiative is its attempt to close it with standards rather than a single vendor's dashboard.

The evidence gap#

Microsoft's framing is blunt: an agent with access to data, tools and credentials carries the risk profile of a privileged insider, but operates at machine speed and scale. When something consequential happens, accountability still sits with the organization — leaders answer to customers, employees, boards, auditors and regulators — and, as Microsoft puts it, blaming the AI is not an answer, and “trust the model” is not a control.

“For security controls to be effective, every organization needs greater transparency into the model itself,” said Hayete Gallot, executive vice president of Microsoft Security. “With this work, we will bring more transparency to all models. This is a critical step in making containment smarter and stronger, allowing us to better inform accountability and traceability.”

The gap matters most after the fact. When an agent moves money, changes a record or grants access, a security team needs to reconstruct what happened — and today much of that reconstruction rests on the model's own logs. An agent should not be the only witness to its own actions.

See, verify, control#

The initiative organizes its answer into three words: see, verify, control. See means knowing what the AI was asked and authorized to do, which model and runtime ran it, what data and tools it touched, and what actions it took — with activity monitored continuously against policies. Verify means the evidence identifies where it came from — the model, the runtime, or independent identity and policy systems — so model-reported signals can be checked against records the model did not write. Control means clear evidence of each workflow's purpose, owner, data and delegated authority, so organizations can enforce policy before a consequential action and reconstruct events during investigations and audits.

Illustration of a security operations desk with monitors showing a glowing trail of connected nodes and verification checkmarks
An agent's activity trail, checked against independent evidence: the “verify” half of Microsoft's framework. Graphic: AI Frontier Post.

One design choice runs through all three: the evidence is supposed to be portable. It should stay with the organization when models and providers change, so a team can swap the model powering an agent without losing the ability to govern or audit it. The initiative is framed as provider-neutral — covering closed and open-weight models alike — and customers should be able to feed the evidence into their own security and compliance tools, Microsoft's or anyone else's.

Standards, not a Microsoft spec#

The announcement itself has three parts. First, Microsoft will expand its participation in the National Institute of Standards and Technology's SI Agent Standards Initiative, drawing on NIST's research to develop open, industry-wide standards for AI transparency, security and governance. Second, it is working with more than 40 organizations — Accenture, Adobe, KPMG, Mastercard and Volvo Cars among them — to define and apply those standards. Third, it will build the resulting work into its own products: Microsoft AI models, Microsoft Foundry and Copilot.

The partner list is doing real work in the pitch: a standard written by one vendor is a product spec, while one written by a consultancy, an auditor, a payments network and a carmaker alongside the vendor is harder to dismiss as marketing. KPMG's global chief information security officer, John Israel, said clients want to move faster with AI without choosing between innovation and safety; Adobe's chief security officer, Aanchal Gupta, added that security teams need clear visibility into what agents access, what they do, and how those actions can be verified.

Illustration of enterprise leaders and engineers gathered around a glass table reviewing a holographic dashboard of AI system activity
More than 40 organizations, from consultancies to carmakers, are collaborating on the standards. Graphic: AI Frontier Post.

The announcement also lands as the industry converges on the same problem from opposite directions. A day earlier, Google's threat-intelligence group reported that AI agents are now turning security patches into working exploits with no human in the loop — and that a North Korea-linked crew had rigged Claude Code's settings to plant a backdoor. Google's report is about what agents can do to you; Microsoft's initiative is about what you can prove they did. Both point the same way: oversight that depends on trusting the agent is no longer enough. Microsoft has also pointed to compliance frameworks such as the EU AI Act as part of the rationale.

What to watch#

First, the hard part: there is no timeline and no dated deliverable. Standards work is slow, consensus is fragile, and 40-plus organizations with competing interests agreeing on common evidence formats is a much bigger test than a keynote slide. The practical test is whether the framework becomes concrete, portable records — or stays a set of principles.

Second, who else signs on. The value of a transparency standard grows with the number of model providers that adopt it, and the announcement names no other frontier lab as a participant. Watch whether OpenAI, Anthropic, Google or the open-weight ecosystem join the NIST process. Third, what shows up first in the products: the first visible implementation in Foundry or Copilot will say more about Microsoft's seriousness than the partner count. The push arrived alongside a broader security package at Ignite, including the general availability of Microsoft Perception — a sign Microsoft is treating agent governance as a product line, not a press release. Organizations interested in shaping the initiative can sign up through Microsoft's site.

Sources: Microsoft Source (November 17, 2026); Microsoft official blog, “Ignite 2026: Building the frontier ecosystem” (November 17, 2026); Microsoft Security blog (November 17, 2026); Unite.AI (November 17, 2026).