Who this is for: Enterprise compliance teams evaluating the AI Transparency Initiative, CISOs governing multi-provider AI estates, auditors and assessors building evidence frameworks for agentic AI, GRC platform architects integrating portable compliance evidence, and organizations among the 40+ initiative partners seeking implementation pathways.
Emerging initiative. Not a final standard. Microsoft announced the AI Transparency Initiative on November 17, 2026, at Ignite in San Francisco. It is a commitment by 40+ organizations to develop open standards, not a published specification. NIST's AI Agent Interoperability Profile is expected in Q4 2026 or early 2027. This guide maps the initiative's stated requirements to evidence procedures that are available today. It will be updated as published standards emerge.
Contents
1. What Microsoft Announced 2. The Evidence Gap 3. SEE / VERIFY / CONTROL Framework 4. SEE: Monitoring and Observability 5. VERIFY: Independent Verification 6. CONTROL: Governance and Authority 7. Adjacent Infrastructure 8. NIST Agent Standards Alignment 9. Quick Reference Table 10. Getting Started 11. Related Guides 12. Scope and Limitations 13. References1. What Microsoft Announced
At Microsoft Ignite 2026 on November 17, CEO Satya Nadella introduced the AI Transparency Initiative alongside more than 40 organizations, including Accenture, Adobe, KPMG, Mastercard, and Volvo Cars. The initiative is organized around a governing principle: greater AI autonomy requires greater transparency.
Nadella framed the challenge in architectural terms. AI systems are becoming more autonomous, making decisions that affect people, infrastructure, and capital. The organizations deploying these systems need to see what their agents are doing, verify that behavior independently, and retain control over what agents are permitted to do. These three capabilities form the initiative's operational framework.
The initiative has three components:
- NIST collaboration: Expand participation in the National Institute of Standards and Technology's AI Agent Standards Initiative (CAISI) to develop open, industry-wide standards for AI transparency, security, and governance.
- Industry partnership: Collaborate with 40+ organizations to define standards for AI systems that are accountable, independently verifiable, and safe by design, regardless of model or provider.
- Product integration: Incorporate the resulting standards into Microsoft AI, Microsoft Foundry, and Microsoft Copilot, so customers can produce evidence compatible with their existing security and compliance tools.
Separately, Microsoft announced Agent 365, a control plane for agent governance, identity, and security, with over 50 million agents under management. Agents connected to Agent 365 automatically receive a Microsoft Entra Agent ID, bringing them into the enterprise identity fabric alongside human users.
Microsoft also announced general availability of Perception, a multi-agent security system where red, blue, and green AI agents run playbooks to simulate attacks, detect threats, and remediate issues.
2. The Evidence Gap
Microsoft identified a structural problem in current AI deployments that the initiative is designed to address. As Nadella stated at Ignite:
Existing security and observability tools can monitor the activity around AI models: what data was sent, what API was called, what response was returned. They cannot reconstruct what happened between a request and an outcome. When a model processes a prompt, reasons about it, and produces a response, the internal path is opaque.
This creates what the initiative calls the evidence gap: the space between what an AI system did and what an organization can independently prove it did. The gap becomes more consequential as AI agents gain autonomy. An agent that calls tools, delegates to sub-agents, and takes sequential actions without human intervention produces outcomes that are increasingly difficult to attribute, audit, or reconstruct.
The initiative's response to the evidence gap is a framework organized around three capabilities: See what agents are doing, Verify that behavior independently, and Control what agents are permitted to do.
Why Independent Evidence Matters
Hayete Gallot, Microsoft's EVP of Security, noted: "For security controls to be effective, every organization needs greater transparency into the model itself." The initiative's verify principle states that agents cannot be the sole source of truth about their own behavior. Organizations should be able to verify AI activity using their own security and compliance systems, without relying solely on the model provider's account.
This principle applies to any provider. Whether the model is hosted on Azure, AWS, GCP, or running locally, the evidence of what that model did should be verifiable by the deploying organization through infrastructure the provider does not control. That is the distinction between platform telemetry and independent evidence.
What the Evidence Gap Looks Like in Practice
Consider a procurement agent deployed on Microsoft Foundry. The agent receives a purchase request, queries a supplier database via an MCP tool hosted on AWS, delegates a credit check to a sub-agent running a third-party model, receives the result, and approves the order. Five steps, three providers, two models, one outcome.
When an auditor asks "why was this supplier selected over the lower-cost alternative?", the evidence trail splits:
- Azure logs show the orchestrator received the request and returned an approval. They do not show what happened inside the MCP tool call or what the sub-agent's credit model evaluated.
- AWS logs show the supplier database query and response. They do not show why the orchestrator chose those query parameters or what it did with the results.
- The third-party model provider shows the credit check was run. It does not show who delegated the check, what authority the sub-agent had, or whether the result was used as-is or modified.
Three platforms, three partial stories, no single party that can reconstruct the full decision chain. This is the evidence gap in practice. Each provider's telemetry is accurate within its own boundary and blind beyond it.
The consequence is real. The auditor writes a finding: "unable to reconstruct decision chain for automated procurement actions." The CISO cannot demonstrate to the board that agent behavior was governed across the full workflow. The organization's EU AI Act transparency obligation (Article 13) is unmet because the evidence exists in fragments that no single authority can assemble. The technical systems all worked correctly. The evidence infrastructure did not.
Now add an independent witness layer. Each step produces a Witness Anchor: the orchestrator's delegation (AI-DEL.1), the tool call to the supplier database (AI-TOOL.1), the sub-agent's credit check (AI-CHAIN.1), the access control decision (AI-ACC.1), and the final approval (AI-HITL.1 if human-reviewed, or AI-AUTO.1 if automated). All five anchors share a lifecycle chain ID. The auditor queries that chain ID and reconstructs the full decision path in one view, verified against the fingerprints from each step, independent of any single provider's account.
The platform telemetry tells you what each system saw. The witness evidence tells you what actually happened across all of them.
Multi-provider agent workflows are where the evidence gap has the most audit consequence. When assessing organizations running agents across Azure, AWS, GCP, or hybrid environments, ask for the lifecycle chain reconstruction. If the organization can only produce platform-specific logs and cannot link steps across provider boundaries, that is a chain-of-custody gap regardless of how complete each provider's individual telemetry is.
3. SEE / VERIFY / CONTROL Framework
The three pillars define what organizations need. SWT3 witness procedures are one way to produce the evidence those pillars require. The mapping below shows where existing procedures align with each pillar's stated capabilities.
| Pillar | Microsoft's Description | Evidence Requirement | SWT3 Procedures |
|---|---|---|---|
| SEE | Monitor what AI was authorized to do, which model or runtime was used, what data and tools were accessed, and what actions were taken | Agent identity, model tracking, action logging, data access records | 37 procedures |
| VERIFY | Obtain evidence covering agent identity, delegated intent, and actions from multiple sources so model-reported signals can be checked against independently observed records | Independent verification, provenance chains, audit integrity, drift detection | 25 procedures |
| CONTROL | Retain authority over data access and workflow governance regardless of model provider | Governance policies, access boundaries, human oversight, guardrails | 32 procedures |
94 procedures across the AI, NHI (Non-Human Identity), HBOM, DPP, and ADR namespaces map to the initiative's three pillars. Each procedure produces a cryptographic Witness Anchor: a provider-neutral, independently verifiable evidence record.
Anatomy of a Witness Anchor
Every SWT3 procedure produces the same evidence artifact: a Witness Anchor. Understanding what an anchor contains makes the pillar mapping concrete. Here is a decomposition of a single anchor from an agent tool call:
SWT3-E-AZURE-AI-TOOL1-PASS-1732000000-a7c3f91b02e8SEE components:
•
agent_id: procurement-agent-01 (which agent acted)•
procedure_id: AI-TOOL.1 (what type of action was witnessed)•
model_id: gpt-4o-2024-11-20 (which model was active)•
tool_name: supplier_search (what tool was called)VERIFY components:
•
fingerprint: a7c3f91b02e8 (SHA-256 digest of tenant + procedure + factors + timestamp, verifiable by any party)•
chain_id: proc-9a4f2c (lifecycle chain linking this anchor to the delegation, credit check, and approval anchors)•
timestamp: 1732000000 (epoch, externally attestable via AI-AUDIT.2)CONTROL components:
•
authorization_id: auth-7721 (pre-authorized scope under which this tool call was permitted)•
clearing_level: 1 (Standard)•
verdict: PASS (the tool call was authorized and completed within policy)
One anchor, three pillars of evidence. The assessor can verify the fingerprint independently (SHA-256 of known inputs), trace the chain_id to reconstruct the full workflow, and confirm the authorization_id against the organization's access control policy. No platform access required.
Privacy-Preserving Verification: Clearing Levels
A common objection to independent evidence is data sensitivity. Regulated organizations -- financial institutions handling trading signals, healthcare systems processing patient data, defense contractors working with classified inputs -- cannot expose raw inference factors to an external audit trail.
The SWT3 clearing engine addresses this by design. Each anchor is minted at a clearing level that controls how much detail survives:
- Level 0 (Analytics): Full factors retained. Suitable for internal observability.
- Level 1 (Standard): Factors retained. Suitable for most compliance evidence.
- Level 2 (Sensitive): Factors redacted, factor hashes retained. The assessor can verify that the factors existed and were consistent, without seeing the values.
- Level 3 (Classified): Only the fingerprint survives. The anchor proves the event occurred and the outcome was recorded, but the inputs are irrecoverable. Even with full database access, the original factors cannot be reconstructed.
At every clearing level, the fingerprint is independently verifiable. A Level 3 anchor proves that an inference happened, that it was witnessed, and that the witness record has not been tampered with. It does not reveal what the model was asked or what it answered. This is how organizations produce compliance evidence for the VERIFY pillar without compromising data classification requirements.
Clearing levels are set per tenant and per procedure, not per anchor. Verify that the clearing level is appropriate for the data classification of the AI workload. Level 0 or 1 for a workload processing classified data is itself a finding. Conversely, Level 3 for a low-sensitivity workload may indicate over-classification that reduces audit utility. The clearing level should match the data's sensitivity, not the organization's default.
When reviewing witness evidence for the first time, start with the chain_id. Pull all anchors sharing that chain and verify the sequence is complete. Then spot-check individual fingerprints against the factor values. If the fingerprint does not match the recomputed hash, the anchor has been tampered with. That is the most fundamental integrity finding.
4. SEE: Monitoring and Observability
SEE: Know What Your Agents Are Doing
37 procedures. Agent identity, model tracking, action logging, data access, credential lifecycle.
The SEE pillar asks organizations to maintain continuous visibility into what their AI systems are authorized to do, which models and runtimes are in use, what data and tools agents access, and what actions they take. This is the observability foundation on which verification and control depend.
Agent Identity
Agent Identity Assertion
Initiative context: Microsoft's Agent 365 assigns every connected agent a Microsoft Entra Agent ID, creating a shared identity inventory for agents and humans. The AI Transparency Initiative extends this principle across providers: every autonomous entity should carry a verifiable, persistent identity.
Evidence produced: Every witness call includes an agent_id field, binding a verifiable identity to every action the agent takes. The identity record is embedded in every Witness Anchor, creating an immutable association between agent and action regardless of which platform hosts the agent.
AI-ID.1 is the identity root for all other agent procedures. Confirm that agent_id values are unique, consistent across sessions, and not shared between unrelated agents. When evaluating Agent 365 deployments, verify that the Entra Agent ID and the witness agent_id are correlated so identity can be traced across both the platform layer and the independent evidence layer.
Credential Scope Attestation
Initiative context: Agent 365 governance capabilities provide granular policies controlling what agents can access. Credential scoping is fundamental to least-privilege enforcement for non-human identities.
Evidence produced: Records the scope assigned to agent credentials, including permitted resources, actions, and boundaries. Produces evidence that credential grants follow the principle of least privilege.
Cross-reference NHI-SCOPE.1 anchors with Entra Agent ID permission grants. Credential scope should match the agent's authorized function. Over-permissioned credentials in a multi-agent environment are a privilege escalation vector identified in NIST red team research.
Credential Lifecycle and Agent-to-Agent Delegation
Initiative context: As agents delegate tasks to other agents, credential chains must be tracked. The initiative specifies that evidence should identify whether a signal comes from the model, the runtime, or independent systems such as identity and policy.
Evidence produced: NHI-CYCLE.1 witnesses credential creation, rotation, and revocation events. NHI-AGENT.1 witnesses agent-to-agent credential delegation, recording which agent delegated credentials to which sub-agent and under what authority.
In multi-agent deployments using Agent 365, verify that credential lifecycle events are witnessed independently of the platform that issued the credentials. Self-reported credential records from the issuing platform are telemetry. Independent witness records are evidence.
Model Tracking
Model Weight Integrity, Version Tracking, Weight File Integrity, and Registry Attestation
Initiative context: The initiative's model-agnostic design requires that the evidence layer works across "closed and open-weight models, so developer teams can change or update the models powering their agents without compromising IT and security teams' ability to govern them securely." Microsoft launched seven new MAI models at Ignite alongside support for Meta and other third-party models in Foundry.
Evidence produced: AI-MDL.1 verifies that the deployed model hash matches the approved hash. AI-MDL.2 records which model version was active at each inference. AI-MDL.5 witnesses weight file integrity for open-weight deployments. AI-MDL.8 attests the model registry state, confirming the model was drawn from an approved catalog.
Model tracking is foundational when organizations use multiple providers. Confirm that AI-MDL.2 records are present for every inference period and that model version changes correlate with approved change records. For Foundry deployments using third-party models, AI-MDL.8 registry attestation is especially relevant because the model lifecycle is managed outside the deploying organization.
Action Logging
Tool Call Witnessing and Tool Permission Attestation
Initiative context: The SEE pillar requires organizations to know what tools agents accessed and what actions they took. Microsoft Perception monitors agent behavior continuously and investigates anomalous actions. The evidence layer needs to record what was authorized as well as what was observed.
Evidence produced: AI-TOOL.1 records every tool invocation: what tool was called, what arguments were provided, what was returned. AI-TOOL.2 attests the permission state of tools at the time they were invoked, distinguishing between authorized and unauthorized tool access.
Tool call witnessing is distinct from tool call logging. Logging records that a tool was called. Witnessing produces a cryptographic receipt that an independent party can verify. For environments using Microsoft Perception alongside SWT3, the Perception record shows threat detection context while the witness anchor shows the compliance evidence chain. Both are valuable; they serve different audit purposes.
Log Retention Compliance
Initiative context: The SEE pillar assumes the existence of structured, retrievable event logs. AI-LOG.1 witnesses the logging infrastructure itself, confirming that logs exist, meet retention requirements, and are accessible for audit.
Evidence produced: Attests that AI operational logs meet minimum retention periods (default: 180 days) and are stored in a retrievable format.
AI-LOG.1 is infrastructure-level evidence. Gaps in AI-LOG.1 during periods of active agent operations indicate logging failures. This is the most fundamental observability finding because it means other SEE-pillar evidence may be incomplete for the gap period.
Human Oversight Tracking
Human Review Completion, Override Event Tracking, and Overseer Identity Capture
Initiative context: Nadella's first pillar is keeping human ambition at the center. Agent 365 provides human approval options to control what agents can do. The evidence layer needs to record when humans reviewed, overrode, or approved agent decisions.
Evidence produced: AI-HITL.1 records human review completion for decisions that require it. AI-HITL.2 witnesses human override events, documenting when and why a human countermanded an agent decision. AI-HITL.3 captures the identity of the reviewing human, creating personal accountability for oversight decisions.
Under the EU AI Act, high-risk AI systems require documented human oversight (Article 14). AI-HITL.1/2/3 anchors provide the evidence chain. Verify that high-risk decisions have corresponding AI-HITL.1 records and that AI-HITL.3 identifies a named, authorized individual, not a service account or generic role.
Data Access
Training Data Provenance, License Compliance, Statistics, PII Lifecycle, and Synthetic Data Provenance
Initiative context: The SEE pillar requires knowledge of what data agents accessed. For model training and fine-tuning, this extends to provenance of the training data itself. Microsoft's 2026 Responsible AI Transparency Report restructured governance around the AI technology stack: models, platform services, and applications.
Evidence produced: AI-DATA.1 records training data sources and lineage. AI-DATA.2 attests license compliance for training datasets. AI-DATA.3 captures statistical characteristics. AI-DATA.4 witnesses PII handling throughout the data lifecycle. AI-DATA.5 records synthetic data provenance, distinguishing generated data from real-world observations.
Data provenance evidence is essential for organizations using fine-tuned or custom models in Foundry. Verify AI-DATA.1 records for any model that was fine-tuned on proprietary data. For models consuming personal data, confirm AI-DATA.4 records exist and align with the organization's data processing agreements.
5. VERIFY: Independent Verification
VERIFY: Confirm Behavior Without Relying on Self-Reporting
25 procedures. Trust verification, evidence provenance, audit integrity, chain of custody, drift detection.
The VERIFY pillar addresses the core challenge the initiative was created to solve. As stated in the announcement: "Agents cannot be the sole source of truth about their own behavior." Organizations should be able to independently verify AI activity using their own security and compliance systems, without depending solely on the model provider's account.
The initiative further specifies that evidence covering agent identity, delegated intent, and actions should identify its source: the model, the runtime around it, or independent systems such as identity and policy. Model-reported signals should be checked against independently observed records.
Trust Verification
Trust Verification, Credential Presentation, and Trust Propagation
Initiative context: The initiative calls for AI systems that are "accountable, independently verifiable and safe by design, regardless of the model or provider." Trust must be established through evidence, not assumed through platform membership.
Evidence produced: AI-TRUST.1 witnesses trust verification events between agents. AI-TRUST.2 records credential presentations during inter-agent trust establishment. AI-TRUST.3 tracks how trust propagates through multi-agent systems, documenting whether trust is direct or inherited.
Trust propagation is a critical finding area in multi-agent deployments. Verify that trust chains are bounded: agent A trusting agent B does not automatically grant agent C the same trust level unless explicitly authorized. Unbounded trust propagation is functionally equivalent to no access control.
Portable Evidence: W3C Verifiable Credentials
The initiative's practical test is whether participants can produce "portable evidence" that works across providers. SWT3 supports exporting compliance evidence as W3C Verifiable Credentials -- a standards-based format that any party can verify without contacting the issuing organization.
- Issuer DID:
did:web(decentralized identifier, resolvable via standard HTTPS) - Proof suite:
eddsa-jcs-2022(Ed25519 digital signatures over JSON Canonicalization Scheme) - Verification: Any party with the public key (published in the DID document) can verify the credential. No API call, no account, no platform dependency.
- Revocation: StatusList2021Entry enables credential revocation without breaking existing verifications.
When an auditor, regulator, or trading partner receives a Verifiable Credential, they verify it against the public key in the DID document. The verification is mathematical, not institutional. This is what "independently verifiable" means in its strongest form: the evidence is self-proving.
W3C Verifiable Credentials are an established standard (W3C Recommendation, 2022). For organizations exchanging compliance evidence across organizational boundaries -- between initiative partners, with regulators, or in vendor assessments -- VCs provide a verification method that does not require the verifier to trust or access the issuer's infrastructure. Ask whether the organization's evidence exports support VC format for cross-boundary scenarios.
Audit Integrity
Audit Log Integrity and External Timestamp Attestation
Initiative context: The initiative's practical test, as described by observers, is whether participants can produce "common standards and portable evidence: records that let an enterprise govern agents before they act and reconstruct their behavior afterward, even when the underlying model or provider changes." Evidence integrity is the foundation.
Evidence produced: AI-AUDIT.1 witnesses audit log integrity, confirming that operational records are tamper-evident. AI-AUDIT.2 provides external timestamp attestation, anchoring events to a verifiable timeline independent of the system being audited.
AI-AUDIT.1 anchors prove the evidence itself is trustworthy. During assessment, verify that audit log hashes are computed over the full log content and that the hash algorithm is deterministic across replay. For cross-provider environments, AI-AUDIT.2 timestamps from an external source (not the model provider) are essential for correlating events across platforms.
Chain of Custody
Multi-Agent Chain Handoff and Chain Trust Degradation
Initiative context: As agents delegate to sub-agents, the chain of custody for decisions and data must remain intact. The initiative requires that organizations can "reconstruct their behavior afterward, even when the underlying model or provider changes."
Evidence produced: AI-CHAIN.1 links multiple anchors via shared lifecycle identifiers, creating a verifiable sequence of events across multi-agent workflows. AI-CHAIN.2 witnesses trust degradation across chains, flagging when downstream agents have weaker verification than upstream agents.
Chain handoff evidence is critical when agents span multiple providers (for example, an orchestrator on Azure delegating to a tool agent on AWS). Verify that lifecycle chain identifiers are consistent across provider boundaries. Missing chain links indicate evidence gaps in the multi-provider workflow.
Delegation Tree Witnessing, Boundary Attestation, and Chain Integrity
Initiative context: Agent delegation is the mechanism by which a simple request becomes a complex multi-step workflow. The NCCoE concept paper identifies access delegation as one of four focus areas for agent governance.
Evidence produced: AI-DEL.1 records delegation events linking parent and child agents. AI-DEL.2 attests the boundaries of delegation authority. AI-DEL.3 witnesses the integrity of the full delegation chain, detecting breaks or unauthorized insertions.
Delegation trees are where privilege escalation occurs in multi-agent systems. Verify that every sub-agent action has a corresponding parent delegation anchor. Orphaned agent actions in a delegation-based architecture indicate untracked delegation, which NIST identifies as a primary attack vector.
Provenance and Supply Chain
Model Provenance Chain, Context Retrieval Provenance, and AI Bill of Materials
Initiative context: The initiative requires portable evidence that follows the model, not the provider. When a model is updated, swapped, or migrated, the provenance chain should remain intact.
Evidence produced: AI-PROV.1 records model provenance from training through deployment. AI-RAG.1 witnesses retrieval-augmented generation context, documenting what documents were retrieved and used in producing outputs. AI-SBOM.1 produces an AI-specific bill of materials, cataloging models, adapters, datasets, and tools in the deployment.
AI-SBOM.1 is the foundational inventory procedure for multi-model environments. Confirm that the SBOM is current and includes all models in production, including third-party models from Foundry marketplace. Organizations using RAG architectures should have AI-RAG.1 records for every retrieval event associated with high-impact outputs.
Drift Detection
Model Drift Detection, Consequence-Mapped Drift Thresholds, and Champion-Challenger Assessment
Initiative context: Microsoft Perception deploys red, blue, and green agents to continuously assess AI security posture. The evidence layer needs to record when model behavior drifts from established baselines and what the consequences of that drift are.
Evidence produced: AI-DRIFT.1 detects statistical drift in model outputs over time. AI-DRIFT.2 maps detected drift to consequence categories, escalating based on impact. AI-ASSESS.1 witnesses champion-challenger model comparisons, documenting performance differences between candidate and production models.
Drift detection evidence is continuous, not periodic. Verify that AI-DRIFT.1 records are present across the full operational period. For organizations also using Perception, the drift evidence from the independent witness layer should corroborate or challenge the platform's own assessment. Discrepancies between self-reported and independently observed drift are significant findings.
6. CONTROL: Governance and Authority
CONTROL: Retain Authority Regardless of Provider
32 procedures. Governance policies, access boundaries, human oversight, guardrails, MCP security, emergency procedures.
The CONTROL pillar requires that organizations retain authority over data access and workflow governance regardless of which model or provider powers their agents. Nadella framed this as an architectural imperative: "It becomes an architectural imperative to separate intelligence from the authority over it."
New Agent 365 governance capabilities provide granular policies and human approval options. Windows Execution Containers enforce containment at the OS level. The evidence layer records that these controls were in effect and functioning.
Governance Framework
Governance Policy Suite
Initiative context: The initiative's 40+ partner organizations will collectively define governance standards. Each participating organization needs evidence that its own governance framework is implemented and enforced.
Evidence produced: AI-GOV.1 attests acceptable use policy existence and currency. AI-GOV.2 records employee AI training completion. AI-GOV.3 witnesses the approved model registry. AI-GOV.4 documents shadow AI incident response. AI-GOV.5 records third-party AI vendor assessments. AI-GOV.6 defines AI risk management scope. AI-GOV.7 attests governance resource allocation.
AI-GOV.3 (Approved Model Registry) is especially relevant for organizations using Foundry with third-party models. Confirm that every model in production appears in the approved registry and that the registry is reviewed on a defined cadence. Shadow AI (AI-GOV.4) incidents involving unauthorized model use should trigger both the organization's incident response and an independent witness record.
Governance Infrastructure Suite
Initiative context: The initiative asks participants to build governance that works across providers. Meta-governance procedures witness the governance infrastructure itself, ensuring the oversight layer is functioning and has not degraded.
Evidence produced: AI-METAGOV.1 attests governance infrastructure health. AI-METAGOV.2 records governance layer registration. AI-METAGOV.3 detects policy downgrades. AI-METAGOV.4 checks for circular dependencies in governance rules. AI-METAGOV.5 attests governance authorization. AI-METAGOV.6 witnesses emergency override attestations. AI-METAGOV.7 verifies governance sync across distributed systems. AI-METAGOV.8 verifies attestation purity (that governance records have not been contaminated by the systems they govern).
AI-METAGOV.3 (Policy Downgrade Detection) is the canary. If governance policies were weakened, whether intentionally or through misconfiguration, the evidence should reflect it. This is distinct from policy violation. A violation breaks an existing policy. A downgrade changes the policy itself. Both require evidence, but they are different findings.
Access Governance
Agent Access Control Witnessing
Initiative context: Agent 365 and Entra provide the enforcement mechanism for access control. The evidence layer records access decisions: what resource was accessed, what permission was evaluated, and whether access was granted or denied.
Evidence produced: Witnesses access control decisions at the agent level, producing cryptographic receipts of authorization evaluations.
AI-ACC.1 anchors should include an authorization_id linking each access decision to a pre-authorized scope. Access granted without a corresponding authorization_id indicates access without pre-authorization, a privilege escalation indicator.
Model Context Protocol Security Suite
Initiative context: As agents connect to external tools and services via protocols like MCP, the security posture of those connections needs evidence. Agent 365 policies are enforced per device; MCP security procedures provide evidence at the protocol layer.
Evidence produced: AI-MCP.1 attests MCP security posture. AI-MCP.2 witnesses tool integrity. AI-MCP.3 records server authentication. AI-MCP.4 attests server discovery processes. AI-MCP.5 witnesses OAuth token binding. AI-MCP.6 witnesses elicitation consent for agent-to-human data requests.
MCP is an emerging standard for agent-to-tool communication. For organizations deploying agents that use external MCP servers, verify that AI-MCP.3 (Server Authentication) records exist for every connected server. Unauthenticated MCP connections are equivalent to unsigned API calls: functional but unverifiable.
Guardrails and Safety
Guardrail Enforcement, Content Safety Filter, and PII Redaction
Initiative context: The initiative calls for AI systems that are "safe by design." Safety guardrails must not only be present but demonstrably active during operation. Microsoft's new MAI Cyber 2 Flash model addresses security-specific use cases; the evidence layer records whether safety filters were engaged across all models.
Evidence produced: AI-GRD.1 attests that required safety filters were active during inference. AI-GRD.2 records content safety classification outcomes. AI-GRD.3 witnesses PII redaction events, confirming that personally identifiable information was handled according to policy.
Guardrail evidence must be continuous, not sampled. Verify that AI-GRD.1 records cover the full operational period with no gaps. A gap in guardrail attestation during active inference indicates either a monitoring failure or a period of unprotected operation. Both are findings.
Emergency Override Lifecycle Witnessing
Initiative context: Microsoft Perception's Guardian Agents can investigate risk changes and recommend containment steps. When an emergency override occurs, whether triggered by Perception, a human operator, or policy, the lifecycle of that override needs evidence from initiation through resolution.
Evidence produced: Witnesses the full lifecycle of emergency overrides: who initiated it, what was overridden, what the justification was, what the resolution was, and when normal operations resumed.
Emergency overrides bypass normal governance. That is their purpose. The evidence requirement is higher because of it. Verify that every AI-EMRG.1 anchor includes an identified initiator (not a service account), a stated justification, and a resolution timestamp. Open-ended overrides without resolution records are the highest-severity governance finding.
Transparency and Fairness
Transparency Disclosure, Bias Measurement, Bias Audit, and Societal Impact Assessment
Initiative context: KPMG Global CISO John Israel stated: "Trust in AI requires transparency, explainability and responsible oversight." The initiative's transparency commitment extends beyond operational monitoring to include fairness and impact evidence.
Evidence produced: AI-TRANS.1 records transparency disclosures. AI-FAIR.1 measures disparity across protected groups. AI-FAIR.3 witnesses formal bias audits. AI-IMPACT.1 records societal impact assessments, documenting the organization's evaluation of downstream effects.
Transparency and fairness evidence is increasingly required by regulation (EU AI Act Article 13, Colorado SB 26-189, NYC Local Law 144). Verify that AI-FAIR.1 disparity measurements use recognized statistical methods and that AI-TRANS.1 disclosures are publicly accessible where required by applicable law.
7. Adjacent Infrastructure
Several SWT3 namespace areas produce evidence that supports the AI Transparency Initiative's goals beyond the core AI procedures:
| Namespace | Procedures | Relevance |
|---|---|---|
| NHI (Non-Human Identity) | NHI-SCOPE.1, NHI-CYCLE.1, NHI-AGENT.1, NHI-ROTATE.1, NHI-REVOKE.1, NHI-PRIV.1 | Agent identity lifecycle evidence for Entra Agent ID and cross-platform credential management |
| HBOM (Hardware BOM) | HBOM-INV.1, HBOM-LIFE.1, HBOM-SUPPLY.1, HBOM-THERM.1, HBOM-WATER.1, HBOM-PUE.1 | Infrastructure-level evidence for AI compute governance, relevant for datacenter-scale AI deployments and sustainability reporting |
| DPP (Digital Product Passport) | DPP-SOH.1, DPP-CHRG.1, DPP-DEGRAD.1, DPP-EOL.1 | Physical product lifecycle evidence, relevant for AI-governed industrial systems (Volvo Cars is an initiative partner) |
| ADR (Automated Demand Response) | ADR-EVENT.1, ADR-BASE.1, ADR-CURT.1 | Energy infrastructure evidence for AI workload governance during demand response events |
Adjacent namespace procedures are relevant when AI governance intersects with physical or infrastructure governance. For organizations like Volvo Cars (initiative partner), DPP and HBOM procedures may be as important as the core AI procedures because their AI systems govern physical products. Assess the full scope of AI-influenced decisions, not just the AI layer in isolation.
8. NIST Agent Standards Alignment
The AI Transparency Initiative explicitly builds on NIST's AI Agent Standards Initiative, launched by CAISI in February 2026. Microsoft committed to expanding its participation in CAISI to help develop open, industry-wide standards.
The NCCoE concept paper (February 5, 2026) identified four focus areas for agent governance. These map directly to both the initiative's pillars and existing SWT3 procedures:
| NCCoE Focus Area | Initiative Pillar | SWT3 Procedures |
|---|---|---|
| Identification | SEE | AI-ID.1, NHI-SCOPE.1, NHI-AGENT.1 |
| Authorization | CONTROL | AI-ACC.1, AI-MCP.5, NHI-PRIV.1 |
| Access Delegation | VERIFY | AI-DEL.1, AI-DEL.2, AI-DEL.3 |
| Logging | SEE | AI-LOG.1, AI-AUDIT.1, AI-AUDIT.2 |
NIST's finalized agent-specific guidance is expected in 2027. The AI Agent Interoperability Profile is targeted for Q4 2026. Organizations building evidence infrastructure today can use the NCCoE focus areas as a stable foundation while awaiting the final standard.
9. Quick Reference Table
| Pillar | Capability | Procedures |
|---|---|---|
| SEE | Agent identity | AI-ID.1, NHI-SCOPE.1, NHI-CYCLE.1, NHI-AGENT.1 |
| Model tracking | AI-MDL.1, AI-MDL.2, AI-MDL.3, AI-MDL.5, AI-MDL.6, AI-MDL.7, AI-MDL.8, AI-BASE.1 |
|
| Action logging | AI-TOOL.1, AI-TOOL.2, AI-LOG.1, AI-HITL.1, AI-HITL.2, AI-HITL.3, AI-LCM.1 |
|
| Data access | AI-DATA.1, AI-DATA.2, AI-DATA.3, AI-DATA.4, AI-DATA.5, AI-CONSENT.1, AI-DPIA.1 |
|
| Incidents | AI-INCIDENT.1, AI-PMM.1 |
|
| VERIFY | Trust verification | AI-TRUST.1, AI-TRUST.2, AI-TRUST.3 |
| Audit integrity | AI-AUDIT.1, AI-AUDIT.2, AI-REV.1, AI-WATERMARK.1 |
|
| Chain of custody | AI-CHAIN.1, AI-CHAIN.2, AI-DEL.1, AI-DEL.2, AI-DEL.3 |
|
| Provenance and drift | AI-PROV.1, AI-RAG.1, AI-RAG.2, AI-MARK.1, AI-SBOM.1, AI-DRIFT.1, AI-DRIFT.2, AI-ASSESS.1, AI-PERF.1, AI-ROBUST.1, AI-EXPL.1, AI-EXPL.2 |
|
| CONTROL | Governance | AI-GOV.1-.7, AI-METAGOV.1-.8 |
| Access and permissions | AI-ACC.1, AI-MCP.1-.6 |
|
| Guardrails | AI-GRD.1, AI-GRD.2, AI-GRD.3, AI-SEC.1, AI-SEC.2, AI-CYBER.1 |
|
| Human oversight | AI-EMRG.1, AI-AUTO.1, AI-AUTO.2, AI-MULTI.1 |
|
| Transparency and fairness | AI-TRANS.1, AI-FAIR.1, AI-FAIR.2, AI-FAIR.3, AI-IMPACT.1, AI-SUPPLY.1, AI-DUALUSE.1 |
10. Getting Started
For Compliance and Security Teams
If your organization is evaluating the AI Transparency Initiative or preparing for NIST agent governance standards, the following actions do not require engineering effort:
- Inventory your agent estate. Identify every AI agent in production, the model it uses, the provider it runs on, and the tools it accesses. This is the SEE pillar baseline. If you cannot produce this inventory today, that is your first gap.
- Identify multi-provider workflows. Any agent workflow that crosses provider boundaries (Azure to AWS, Foundry to self-hosted, cloud to on-premises) is where the evidence gap is widest. These workflows need independent chain-of-custody evidence first.
- Define your evidence requirements. For each pillar, determine what your organization needs to prove:
- SEE: Can you show which agent acted, which model was used, and what tools were accessed for any given decision?
- VERIFY: Can you reconstruct a multi-step agent workflow from independent evidence, without relying on any single provider's logs?
- CONTROL: Can you demonstrate that governance policies were in effect and enforced during the period under review?
- Request a lifecycle chain reconstruction. Ask your engineering team to produce the full evidence chain for one multi-step agent workflow. If they can, you have a foundation. If they cannot, the evidence gap is real and you now know where it is.
Steps 1 and 2 above are assessment preparation tasks that any organization should complete before an AI governance review. The agent inventory (step 1) maps to AI-SBOM.1. The multi-provider identification (step 2) maps to AI-CHAIN.1. If the organization has not completed these steps, begin the assessment there.
For Engineering Teams
The SWT3 protocol is open-source (Apache 2.0) and available in 12 languages. Integration requires adding a witness layer to your existing agent infrastructure. The witness operates alongside your model client, not in place of it.
# Python
pip install swt3-ai
from swt3_ai import Witness
witness = Witness(
endpoint="https://your-instance.example.com/api/v1/witness",
api_key="axm_your_key",
agent_id="procurement-agent-01" # SEE: AI-ID.1
)
# Wrap your model client for automatic witnessing
client = witness.wrap(your_openai_client) # VERIFY: every inference witnessed
# Tool call witnessing # SEE: AI-TOOL.1
result = witness.wrap_tool(tool_name="search_db", tool_input=query)
# Access control witnessing # CONTROL: AI-ACC.1
witness.wrap_access(resource="customer_records", action="read", granted=True)
# Flush to produce anchors
witness.flush()
// TypeScript
import { Witness } from '@tenova/swt3-ai';
const witness = new Witness({
endpoint: 'https://your-instance.example.com/api/v1/witness',
apiKey: 'axm_your_key',
agentId: 'procurement-agent-01' // SEE: AI-ID.1
});
// Wrap for automatic witnessing
const client = witness.wrap(yourOpenAIClient); // VERIFY: every inference witnessed
// Tool call witnessing // SEE: AI-TOOL.1
const result = await witness.wrapTool('search_db', query);
// Flush to produce anchors
await witness.flush();
SDKs are also available for Rust, C#, Ruby, Go, Kotlin, and Swift. For infrastructure-as-code deployment, see the Terraform/OpenTofu deployment guide.
11. Related Guides
The following guides provide deeper coverage for specific aspects of the AI Transparency Initiative's requirements:
| Topic | Guide | Relevance |
|---|---|---|
| Microsoft platforms | Microsoft AGT Crosswalk | OWASP Agentic Top 10 mapping for Foundry, Scout, and Windows Agent Runtime. How AGT runtime enforcement and independent witnessing work together. |
| Microsoft platforms | Microsoft Foundry Witness Guide | Step-by-step witness integration for Foundry Agent Service deployments. |
| NIST standards | NIST Agentic AI Standards Crosswalk | Detailed mapping to CAISI pillars and NCCoE concept paper. The initiative builds directly on this NIST work. |
| NIST standards | NIST AI 600-1 GenAI Profile Crosswalk | Generative AI risk profile procedures. Relevant for Foundry and Copilot deployments using generative models. |
| Multi-agent evidence | Multi-Agent Chain Audit Trail | How lifecycle chains link agent actions across delegation boundaries. The chain reconstruction described in the evidence gap scenario. |
| Multi-agent evidence | Trust Mesh for Multi-Agent Systems | Trust propagation, verification, and degradation across agent networks. Maps to AI-TRUST.1/2/3. |
| Multi-agent evidence | Multi-Agent Phased Witnessing | Phased workflows with multiple agents as milestones. Loan underwriting scenario with SR 26-2 scope. |
| Infrastructure | Terraform/OpenTofu IaC Deployment | Infrastructure-as-code modules for AWS Lambda, ECS, EKS, Azure, and GCP. Deploy the witness layer alongside your agent infrastructure. |
| ISO 42001 | ISO 42001 Implementation Evidence | AI management system evidence for certification. Many initiative partners will pursue ISO 42001 alongside the transparency standards. |
| Agent governance | CSA AI Agent Governance Crosswalk | Cloud Security Alliance agent governance mapping. Complementary standards body work alongside NIST CAISI. |
| Agent governance | OWASP Agentic AI Top 10 Crosswalk | The 10 agent-specific attack surfaces. Perception addresses detection; witness evidence addresses accountability. |
| Enterprise readiness | Enterprise Readiness Guide | Production deployment checklist, SLA, security architecture, and compliance posture for enterprise evaluators. |
| EU AI Act | EU AI Act Crosswalk | Article-level mapping for high-risk AI obligations. Several initiative partners (Volvo, Accenture, KPMG) operate under EU jurisdiction. |
12. Scope and Limitations
- Not a final standard. The AI Transparency Initiative is an announced commitment by 40+ organizations. No published specification, control catalog, or compliance checklist exists as of November 2026. This mapping is based on stated requirements from the announcement, the accompanying blog post, and related NIST materials. It will be updated as formal standards emerge.
- Complementary, not exhaustive. SWT3 produces one type of evidence: provider-neutral, cryptographic witness records. Organizations will likely need multiple evidence sources to satisfy the initiative's full scope. Platform telemetry from Azure, AWS, or GCP provides operational context. Security tools like Microsoft Perception provide threat detection. SWT3 provides the independent, verifiable evidence layer.
- Provider-neutral by design. This mapping applies equally to Microsoft, AWS, GCP, open-weight, and hybrid environments. The witness protocol does not privilege any provider. Evidence produced for a Foundry deployment has the same format, verification method, and trust properties as evidence produced for a Bedrock or self-hosted deployment.
- NIST standards are not finalized. NIST's CAISI Agent Standards Initiative and NCCoE concept paper are directional documents. The AI Agent Interoperability Profile is targeted for Q4 2026. Final guidance may alter the mapping in this document.
13. References
- Microsoft AI Transparency Initiative announcement (November 17, 2026)
- Ignite 2026: Building the Frontier Ecosystem (Microsoft Blog, November 17, 2026)
- Security for the Agentic Era, Powered by Microsoft Perception (Microsoft Security Blog, November 17, 2026)
- What's New in Microsoft Agent 365: Ignite 2026 (Microsoft Tech Community)
- 2026 Responsible AI Transparency Report (Microsoft)
- NIST AI Agent Standards Initiative Explained (CASRAI)
- NIST AI Agent Standards: Listening Sessions and Emerging Controls (Cloud Security Alliance)
- NIST Agentic AI Standards Crosswalk (SWT3 Protocol Guide)
- Microsoft Agent Governance Toolkit Crosswalk (SWT3 Protocol Guide)
- CSA AI Agent Governance Crosswalk (SWT3 Protocol Guide)
- UCT Registry (293 procedures, 184 frameworks)
- SWT3 Protocol Specification (swt3.ai)
- Public Anchor Verification (swt3.ai/verify)