Safety & Security

Microsoft’s AI Transparency Initiative Targets Enterprise Agents

Microsoft announced its AI Transparency Initiative at Ignite on November 17, 2026, with more than 40 organizations. The framework seeks portable evidence for agent actions across providers.

Microsoft Launches AI Transparency Initiative for Enterprise Agents
Microsoft Launches AI Transparency Initiative for Enterprise Agentsfrabre / pixabay
By Rebecca Stone5 min read

Why it matters

  • Microsoft announced the AI Transparency Initiative on November 17, 2026, at Microsoft Ignite.
  • More than 40 organizations joined the effort to develop cross-provider open standards.
  • The framework has three parts: See, Verify and Control.
  • Microsoft will expand participation in NIST’s SI Agent Standards Initiative.
  • Microsoft plans to incorporate the work into Microsoft AI, Microsoft Foundry and Microsoft Copilot.

By Rebecca Stone

Microsoft announced the AI Transparency Initiative on November 17, 2026, at Microsoft Ignite, where more than 40 organizations joined an effort to develop open standards for securing and governing AI models and agents across providers.

Chairman and CEO Satya Nadella introduced the initiative with customers and partners including Accenture, Adobe, KPMG, Mastercard and Volvo Cars. Microsoft said the collaboration aims to make AI systems accountable, independently verifiable and safe by design regardless of the model or provider.

Microsoft said enterprises now connect models to data, tools, credentials and infrastructure, allowing agents to take actions without human approval at every step. Microsoft said the governance problem extends beyond whether an agent can be discovered, governed or contained. It also concerns what the agent did and whether the organization can independently verify that account.

Why do current agent controls leave an evidence gap?

Microsoft said existing controls can observe activity around models but cannot see what happens inside them or independently verify models’ accounts of their own behavior. On Microsoft’s Security blog, the company made a similar distinction for agent containment: discovering, governing and containing an agent still does not reveal how its underlying model behaves.

Microsoft said that leaves three questions after consequential activity:

  • What was the AI asked and authorized to do?
  • What happened between the request and the outcome?
  • Can the activity be independently verified?

Microsoft said the resulting evidence gap creates security, compliance, audit and accountability risks. It compared agents with access to data, tools and credentials to “privileged insiders operating at machine speed and scale.”

For corporate leaders, the responsibility remains with the deploying organization. Microsoft said executives must still answer to customers, employees, boards, auditors and regulators when an agent acts on company systems. Microsoft said a model provider’s description of that conduct does not remove the enterprise’s need for evidence.

How does the See, Verify and Control framework work?

Microsoft organized the requirements into a three-part framework named See, Verify and Control. Microsoft said each part connects a different class of evidence to an enterprise’s security and compliance work.

  • See. This layer covers the request and authorization, the model and runtime, the data and tools accessed, and the actions taken. It also calls for continuous monitoring of agent activity and risk signals against policies and thresholds.
  • Verify. This layer addresses the source of evidence. Records of agent identity, delegated intent and actions would show whether information came from the model, its runtime, or independent identity and policy systems.
  • Control. This layer uses evidence of each workflow’s purpose, owner, data and delegated authority to enforce policies before consequential actions. Those records also support reconstruction during investigations, containment, recovery and audits.

Under Verify, model-reported signals could be checked against independently observed records through an organization’s own security and compliance systems. Microsoft said this approach relies on comparison rather than the provider’s account alone.

Control applies evidence before consequential actions and supports review afterward. Microsoft said the same records can support investigations, containment and recovery, as well as compliance audits. The workflow’s purpose, owner, data access and delegated authority provide the basis for those controls.

For an audit, Microsoft said the records would cover the agent’s identity, delegated intent and actions. Control evidence would add the workflow’s purpose, owner, data access and delegated authority. A reviewer could compare what the agent was authorized to do with what the independently observed records show it did.

Why must the evidence remain portable?

Microsoft said the records should remain usable when an organization changes models or providers. The design covers both closed and open-weight models, allowing development teams to update agents without compromising security governance or compliance audits.

Microsoft’s governing principle is that “greater AI autonomy requires greater transparency.” Under the framework, the evidence trail cannot depend solely on what a model or provider reports about its own conduct. Independent identity, policy and observed-action records provide a basis for verification across the system’s lifecycle.

Across the framework, Microsoft ties identity, delegated intent, actions, workflow purpose, ownership, data and delegated authority to the evidence. Together, those fields let an enterprise distinguish what was requested, what the agent was permitted to do, and what it actually did.

How will NIST turn the framework into standards?

Microsoft will expand its participation in NIST’s SI Agent Standards Initiative and draw on its own research to develop open, industry-wide standards for AI transparency, security and governance.

Microsoft said the NIST program is designed to foster industry-led technical standards and open protocols so autonomous agents can operate securely and interoperably across digital systems. The program combines three strategic pillars:

  • Voluntary guidelines and gap analyses for industry-led standards.
  • Work to reduce barriers to interoperable, community-led protocols.
  • Research into agent authentication, identity infrastructure and security evaluations.

Where will Microsoft put the initiative’s work?

Microsoft plans to incorporate the initiative’s work into Microsoft AI, Microsoft Foundry and Microsoft Copilot. Customers would then be able to use the resulting evidence with security and compliance tools from Microsoft or other providers.

That product route follows the framework’s portability requirement. Microsoft said customers could use the evidence with their own security and compliance systems as well as tools from Microsoft or other providers. Model-reported signals would still face comparison with independently observed records.

What must participants prove?

The initiative’s practical test is whether participants can turn the framework into common standards and portable evidence. Microsoft set the target plainly: records must let an enterprise govern agents before they act and reconstruct their behavior afterward, even when the underlying model or provider changes.

That requirement joins pre-action governance to post-action reconstruction. It also makes a provider switch a test of the standards, rather than an exception to them. Development teams can then update agents without compromising security governance or compliance audits.

The next test is whether participants can convert the framework into shared standards and records that remain usable after an enterprise changes models or providers.

Source: unite.ai

Share this article:

More from Rebecca Stone

Rebecca Stone

Show full bio

Correspondent covering companies, deals and strategy in artificial intelligence and machine learning at AI In Context.

728 articles

Related articles

  1. Okta-Led Blueprint Alliance Wants a Kill Switch for Every AI Agent
  2. PwC survey: AI accountability scattered, 11% say it's unclear
  3. 83% of Firms Run Unapproved AI Coding Agents, BlueFlag Research Finds
  4. AI Security in 2027: Agent Permissions Will Define the Blast Radius
  5. Australian firm opens the books on five months of AI agent controls

« Previous articleNext article »