Who this is for: Cloud security architects, GRC teams managing audit evidence, AI platform engineers responsible for governance pipelines, C3PAOs and assessors evaluating AU-2/AU-3 compliance, and DevOps teams operating AI workloads on AWS, Azure, or GCP.

Prerequisite: Basic familiarity with cloud audit logs (CloudTrail, Azure Activity Log, or GCP Cloud Audit Logs) and the SWT3 witness protocol. No prior OCSF knowledge required.

Assessor Notice: SWT3 witness anchors provide timestamped, tamper-evident records of infrastructure events. They supplement -- but do not replace -- native cloud provider audit logs, SIEM retention policies, or assessor judgment. The binary processes audit events locally; raw log data never leaves the customer's environment. Only canonical hashes and human-readable summaries are transmitted to the witness ledger.

Contents

1. The Evidence Gap 2. How Infrastructure Witnessing Works 3. The 6 Canonical Fields 4. Four Cloud Adapters 5. INF-OCSF.1: The Procedure 6. The Unified Merkle Tree 7. Clearing Levels for Infrastructure Evidence 8. Quick Start 9. Three Assessment Scenarios 10. Honest Scope 11. Related Guides and References

1. The Evidence Gap

Every organization running AI in production generates two distinct evidence streams. The first is AI governance evidence: what the model decided, what guardrails fired, whether drift was detected, who authorized the inference. The SWT3 SDK witnesses this evidence today.

The second is infrastructure evidence: what cloud API calls were made, who made them, what resources were created or modified, and whether those calls succeeded. AWS CloudTrail, Azure Activity Log, and GCP Cloud Audit Logs capture this data. Every cloud customer already has these logs.

The problem is that these two evidence streams live in completely separate systems. CloudTrail logs sit in S3 buckets. AI governance anchors sit in the witness ledger. When an assessor asks "was the AI inference that triggered this infrastructure change properly governed?", there is no way to produce a single proof that covers both events. You can show the AI decision. You can show the cloud API call. But you cannot cryptographically prove they coexisted in the same evidence commitment.

This gap matters more every quarter. AI agents are now deploying infrastructure autonomously. Pulumi AI, Terraform Cloud agents, and coding assistants with cloud access make API calls that change your environment. The AI decision and the infrastructure change are two halves of the same governance question. Without a unified evidence stream, assessors must manually correlate timestamps across disparate systems and trust that neither log was tampered with independently.

Assessor Tip

When evaluating AU-2 (Audit Events) and AU-3 (Content of Audit Records) for organizations running AI workloads, ask whether the audit evidence covers both the AI decision layer and the infrastructure execution layer. CloudTrail alone satisfies AU-2 for API calls. SWT3 alone satisfies it for AI inference. Neither alone covers the full decision chain from model output to infrastructure change. The unified tree closes this gap.

2. How Infrastructure Witnessing Works

The swt3-infra-witness binary is a zero-dependency Go program (~9MB static binary) that reads cloud audit log events, extracts a fixed set of canonical fields, computes a deterministic hash, generates a human-readable summary, and sends only the hash and summary to the SWT3 witness ledger. The raw audit data never leaves your environment.

The flow is straightforward:

  1. Your cloud provider generates an audit event (CloudTrail JSON, Azure Activity Log, or GCP Audit Log).
  2. The binary reads the event from a file, stdin pipe, or HTTP webhook.
  3. It auto-detects the format (CloudTrail, Azure, GCP, or pre-normalized OCSF) and extracts exactly 6 canonical fields.
  4. It computes a SHA-256 hash of the canonical fields using deterministic JSON serialization (sorted keys, compact format).
  5. It generates a human-readable summary such as: "At 2026-10-31T14:22:07Z, arn:aws:iam::123:role/deploy called RunInstances on i-0abc123 in us-east-1 [Success]".
  6. It sends the hash and summary to the SWT3 witness ledger, where the evidence is anchored and included in the daily Merkle tree alongside AI inference anchors.

The binary operates in two modes:

ModeUse CaseHow It Works
CLI CI/CD pipelines, cron jobs, ad hoc analysis Reads a file or stdin, processes all events, sends to the ledger, exits. Maintains a cursor file for deduplication across runs.
Server SIEM forwarding, EventBridge rules, continuous monitoring Long-running HTTP receiver. POST audit events to the root path, health check at /healthz. Processes events as they arrive.
Assessor Tip

The binary's architecture preserves data sovereignty by design. Raw CloudTrail/Azure/GCP data stays in the customer's account. Only the canonical hash (a one-way function) and a human-readable summary reach the witness ledger. This means the organization can satisfy AU-3 (Content of Audit Records) locally while the witness ledger provides AU-10 (Non-repudiation) through independent, tamper-evident anchoring. Ask to see the binary's deployment location to confirm it runs inside the customer's trust boundary.

3. The 6 Canonical Fields

Every cloud audit event, regardless of provider, is reduced to exactly 6 fields before hashing. This normalization is what makes the hash cloud-agnostic: the same 6 fields, the same SHA-256 algorithm, the same deterministic serialization, regardless of whether the event came from AWS, Azure, GCP, or a SIEM that already normalized it to OCSF.

FieldWhat It CapturesCloudTrail SourceAzure SourceGCP Source
operation The API call name eventName operationName.value methodName
actor Who made the call userIdentity.arn caller authenticationInfo.principalEmail
resource What was acted on resources[0].ARN resourceId resourceName
time When it happened (UTC) eventTime eventTimestamp timestamp
status Success or error code errorCode or "Success" status.value status.code
region Cloud region awsRegion resourceLocation resource.labels.location

The field set is pinned. It does not change with OCSF schema evolution or cloud provider API updates. This means a hash computed from a CloudTrail event today will be verifiable against the same canonical fields years from now, regardless of how the provider's log format evolves.

The 6 fields are sorted alphabetically by key, serialized to compact JSON (no whitespace, no HTML escaping), and hashed with SHA-256. The result is a 64-character hex string that deterministically identifies the infrastructure event.

Assessor Tip

The pinned field set is a durability guarantee. For AU-12 (Audit Record Generation) assessments, verify that the 6 canonical fields cover the minimum content requirements: what happened (operation), who did it (actor), what was affected (resource), when (time), the outcome (status), and where (region). These map directly to NIST 800-53 AU-3 content elements. The deterministic hash means the same event always produces the same fingerprint, enabling independent verification at any future point.

4. Four Cloud Adapters

The binary ships with four format adapters. Each adapter maps a cloud-specific audit log format to the 6 canonical fields. The format is auto-detected by default, or you can specify it explicitly.

AWS CloudTrail

The default starting point for most organizations

CloudTrail is enabled by default in every AWS account. The adapter reads CloudTrail JSON in either single-event or {"Records": [...]} envelope format. It extracts the 6 canonical fields from CloudTrail's native schema and classifies each event by OCSF activity type (Create, Read, Update, Delete).

Typical sources: CloudTrail S3 delivery, CloudWatch Logs export, EventBridge rules, Security Lake (pre-normalized to OCSF).

Assessor Tip

CloudTrail management events are enabled by default, but data events (S3 object operations, Lambda invocations) require explicit enablement. For a complete AU-2 assessment, confirm which CloudTrail event types are enabled and whether the witness coverage matches the declared audit event scope. Events not captured by CloudTrail cannot be witnessed.

Azure Activity Log

Resource management operations across Azure subscriptions

Azure Activity Log captures control-plane operations at the subscription level: resource creation, policy changes, role assignments, and diagnostic settings. The adapter maps Azure's operation name, caller identity, resource ID, timestamp, status, and location to the 6 canonical fields.

Typical sources: Activity Log export to Storage Account, Event Hub streaming, Azure Monitor diagnostic settings.

Assessor Tip

Azure Activity Log covers the control plane (resource management) but not the data plane (blob reads, database queries). For organizations running AI workloads on Azure, confirm that Activity Log diagnostic settings include all relevant categories (Administrative, Security, Policy, Recommendation). Data-plane auditing requires Azure Diagnostics Logs, which are a separate configuration.

GCP Cloud Audit Logs

Admin Activity, Data Access, and System Event logs

GCP Cloud Audit Logs capture three categories: Admin Activity (always on, free), Data Access (must be enabled, billed), and System Events (always on, free). The adapter maps GCP's method name, principal email, resource name, timestamp, status code, and location to the 6 canonical fields.

Typical sources: Cloud Logging export to Cloud Storage, Pub/Sub topic routing, BigQuery sink.

Assessor Tip

GCP Admin Activity logs are enabled and retained for 400 days by default with no configuration required. Data Access logs must be explicitly enabled per service and incur additional costs. When assessing AU-2 coverage for GCP, verify that Data Access logging is enabled for AI-related services (Vertex AI, Cloud Functions, Cloud Run) if those services are in scope.

OCSF (Pre-Normalized)

For SIEMs that already output OCSF-formatted events

The Open Cybersecurity Schema Framework (OCSF) is a vendor-neutral schema adopted by Amazon Security Lake, Splunk, CrowdStrike, and other security platforms. If your SIEM already normalizes events to OCSF, the binary accepts them directly without re-parsing the cloud-native format. The adapter maps OCSF's api.operation, actor.user.name, dst_endpoint.name, time, status, and cloud.region to the 6 canonical fields.

Typical sources: Amazon Security Lake, any SIEM with OCSF export, custom normalization pipelines.

Assessor Tip

When infrastructure events arrive pre-normalized via OCSF, verify the normalization fidelity. The witness binary re-derives the canonical hash server-side as a tamper detection measure. If a SIEM's OCSF normalization drops or truncates a field, the server-side hash will not match the client-side hash, and the event will be flagged. This is a built-in integrity check, not a bug.

5. INF-OCSF.1: The Procedure

Infrastructure evidence is anchored under a single procedure: INF-OCSF.1 (Infrastructure OCSF Evidence). Each witnessed infrastructure event produces one SWT3 anchor with the procedure ID INF-OCSF.1.

INF-OCSF.1

Infrastructure OCSF Evidence

Records that a cloud infrastructure audit event was captured, normalized to canonical fields, hashed deterministically, and anchored in the witness ledger. The anchor proves that a specific API call occurred at a specific time, made by a specific actor, affecting a specific resource, with a specific outcome. The canonical hash enables independent verification without access to the raw audit log.

Evidence anchored: Canonical hash (SHA-256 of 6 sorted fields), human-readable event summary, OCSF activity type classification (Create/Read/Update/Delete), clearing level applied.

Framework Crosswalks

INF-OCSF.1 maps to audit and accountability controls across 7 frameworks:

FrameworkControlWhat INF-OCSF.1 Satisfies
NIST 800-53AU-2, AU-3, AU-6, AU-12Audit events, content, review/analysis, generation
EU AI ActArt. 12(1)Automatic logging for high-risk AI traceability
FedRAMP HIGH/MODAU-2, AU-3, AU-6, AU-12Same NIST controls, FedRAMP-scoped
CMMC Level 2AU.L2-3.3.1System-level audit logging
NIST 800-1713.3.1Audit log creation and protection
SOC 2CC7.2Monitoring of system components
NIST AI RMFGOVERN 1.7Processes for AI risk governance documentation
Assessor Tip

For FedRAMP assessments, INF-OCSF.1 anchors provide independent corroboration of CloudTrail evidence. The assessor can verify that the canonical hash in the witness ledger matches a re-computation from the raw CloudTrail event in the customer's S3 bucket. If the hashes match, the event has not been tampered with since it was witnessed. This is particularly valuable for AU-10 (Non-repudiation), where the requirement is evidence that an action was taken by a specific entity and cannot be denied.

6. The Unified Merkle Tree

This is the central concept. Every SWT3 anchor -- whether from an AI inference (witnessed by the SDK) or an infrastructure event (witnessed by this binary) -- lands in the same witness ledger and participates in the same daily Merkle root computation.

The daily Merkle root is a single SHA-256 hash that cryptographically commits to every anchor minted that day. A membership proof can verify that any specific anchor was included in the root without revealing the other anchors. This is standard Merkle tree cryptography, the same mechanism used in certificate transparency logs and blockchain systems.

What makes the unified tree unique is cross-domain coverage. A single membership proof can verify that:

were both included in the same daily cryptographic commitment. This is the proof that the AI decision and the infrastructure action coexisted in the same evidence record. No other system on the market produces this cross-domain proof.

Example: Daily Merkle Root

Root: e7a31b...
├ AI-INF.1 -- Model inference witnessed by SDK
├ AI-GRD.1 -- Guardrail evaluation witnessed by SDK
├ INF-OCSF.1 -- RunInstances witnessed by infrastructure binary
├ INF-OCSF.1 -- CreateSecurityGroup witnessed by infrastructure binary
└ Single root covers both evidence domains
Assessor Tip

The unified tree enables a new class of assessment evidence: cross-domain correlation proofs. When evaluating AU-6 (Audit Record Review, Analysis, and Reporting), ask whether the organization can demonstrate that AI decisions and their resulting infrastructure changes are correlated in a single tamper-evident record. The Merkle membership proof provides mathematical certainty that both events were committed together, which is stronger evidence than timestamp correlation across separate log systems.

7. Clearing Levels for Infrastructure Evidence

Infrastructure evidence uses the same 4-tier clearing system as AI inference evidence. The clearing level determines how much human-readable detail is retained in the witness ledger versus cleared before transmission.

LevelNameWhat Is RetainedWhat Is Cleared
0 Analytics Full summary, all 6 fields visible in observations Nothing
1 Standard Full summary, actor identity hashed in observations Raw actor identity (IAM ARN, principal email)
2 Sensitive Operation and status only Actor, resource, region
3 Classified Canonical hash only All human-readable content

Level 1 is the default and appropriate for most organizations. It provides enough detail for assessors to understand what happened while protecting the specific IAM identity that made the call. Level 0 is for non-sensitive environments where full visibility is desired. Level 3 is for classified environments where even the operation name must not leave the local environment.

Regardless of clearing level, the canonical hash is always retained. This means even at Level 3, the organization can independently prove that a specific infrastructure event occurred by re-computing the hash from their local audit log and matching it against the anchored hash in the ledger.

Assessor Tip

The clearing level should align with the data classification of the underlying audit logs. For FedRAMP Moderate systems, Level 1 (Standard) is typically appropriate -- it clears the raw IAM identity while retaining enough operational context for AU-6 review. For systems processing CUI or classified data, Level 2 or Level 3 may be required. Verify that the configured clearing level matches the system's authorization boundary documentation. Even at Level 3, the canonical hash enables independent verification by anyone with access to the raw log.

8. Quick Start

Three steps from zero to witnessed infrastructure evidence.

Step 1: Build the binary

# Zero external dependencies, ~9MB static binary
git clone https://github.com/tenova-labs/swt3-infra-witness
cd swt3-infra-witness
CGO_ENABLED=0 go build -o swt3-infra-witness ./cmd/swt3-infra-witness/

Step 2: Verify with a dry run

# No API key needed -- parse, extract, hash, and preview without sending
./swt3-infra-witness --dry-run --input sample-events/cloudtrail-sample.json

The dry run shows you exactly what will be witnessed: operation, status, canonical hash, fingerprint, and human-readable summary for each event. Nothing is sent to the ledger.

Step 3: Witness for real

# Set your API key and run
export SWT3_API_KEY=axm_your_key
./swt3-infra-witness --input your-cloudtrail-events.json

The binary auto-detects the input format (CloudTrail, Azure, GCP, or OCSF). For explicit format selection:

# Pipe from AWS CLI
aws cloudtrail lookup-events --output json | ./swt3-infra-witness --format cloudtrail

# Azure Activity Log export
./swt3-infra-witness --format azure --input azure-activity-export.json

# GCP Cloud Audit Logs
./swt3-infra-witness --format gcp --input gcp-audit-logs.json

# Server mode for continuous witnessing (SIEM webhook target)
./swt3-infra-witness --mode server --port 8080

Docker

# ~7MB image (distroless base + static binary)
docker run -e SWT3_API_KEY=axm_your_key ghcr.io/tenova-labs/swt3-infra-witness:0.1.0 \
  --mode server --port 8080

Configuration

All configuration is via environment variables (12-factor):

VariableRequiredDefaultDescription
SWT3_API_KEYYes--Your SWT3 API key
SWT3_CLEARING_LEVELNo1Evidence clearing level (0-3)
SWT3_AGENT_IDNoswt3-infra-witnessInstance identity for multi-binary deployments
SWT3_SIGNING_KEYNo--HMAC-SHA256 signing key for anchor authentication
Assessor Tip

During a CM-3 (Configuration Change Control) assessment, verify that the binary is deployed inside the authorization boundary and that the API key is stored in a secrets manager, not hardcoded. The SWT3_AGENT_ID should be unique per deployment so that anchors from different environments (staging vs. production) are distinguishable in the ledger. For HMAC-signed anchors, verify the signing key is rotated according to the organization's key management policy.

9. Three Assessment Scenarios

Scenario A: AI agent deploys infrastructure

A Claude Code session creates an EC2 instance to serve a fine-tuned model. The SWT3 SDK witnesses the model inference (AI-INF.1). CloudTrail captures the RunInstances call. The infrastructure binary witnesses the CloudTrail event (INF-OCSF.1). Both anchors land in the same daily Merkle root.

Assessment value: The assessor can request a membership proof covering both anchors. This proves that the AI decision and the infrastructure deployment were committed in the same cryptographic record. If either event were fabricated after the fact, the Merkle proof would fail verification.

Assessor Tip

For organizations using AI agents with cloud access (Pulumi AI, Terraform Cloud agents, coding assistants), the unified tree provides the only way to prove the AI decision and its infrastructure consequence were both governed. Ask the organization to demonstrate a cross-domain membership proof. If they can produce one, both evidence streams are linked. If they cannot, the AI governance and infrastructure audit trails are independent and must be correlated manually.

Scenario B: FedRAMP authorization with AI workloads

A cloud service provider seeking FedRAMP authorization runs AI inference as part of their service. The 3PAO needs to evaluate AU-2 through AU-12 for both the AI layer and the infrastructure layer. With infrastructure witnessing, the provider can present a single evidence stream where CloudTrail events and AI governance anchors share the same ledger, the same Merkle roots, and the same verification interface.

Assessment value: Instead of presenting CloudTrail exports in one format and AI governance evidence in another, the provider presents a unified ledger. The 3PAO can verify any anchor -- AI or infrastructure -- using the same process. The INF-OCSF.1 crosswalk to AU-2/AU-3/AU-6/AU-12 maps directly to the controls being assessed.

Assessor Tip

For 3PAO assessments, infrastructure witness anchors complement (not replace) CloudTrail as the primary audit evidence. The witness anchor provides non-repudiation (AU-10) and centralized review capability (AU-6) that CloudTrail alone does not offer. The canonical hash enables cross-account verification: the assessor can re-compute the hash from the raw CloudTrail event in the customer's S3 bucket and compare it to the hash in the witness ledger. A match confirms the event has not been altered since it was witnessed.

Scenario C: EU AI Act Art. 12(1) automatic logging

Article 12(1) of the EU AI Act requires high-risk AI systems to have "automatic logging capabilities" that enable traceability. For AI systems deployed on cloud infrastructure, traceability requires knowing not just what the model decided but what infrastructure state existed when the decision was made. Infrastructure witnessing provides the infrastructure half of the Art. 12(1) evidence.

Assessment value: The unified tree demonstrates that the organization's logging covers both the AI decision layer and the infrastructure execution layer. This is a stronger Art. 12(1) compliance posture than logging AI decisions alone.

Assessor Tip

When evaluating Art. 12(1) compliance for high-risk AI systems deployed on cloud infrastructure, the question is whether the "automatic logging" covers the full decision chain. If the AI system's output triggers infrastructure changes (scaling decisions, deployment actions, configuration modifications), those changes are part of the traceability requirement. INF-OCSF.1 anchors provide the infrastructure portion of that chain.

10. Honest Scope

Infrastructure witnessing covers cloud control-plane API calls. It does not cover everything. Being clear about boundaries helps assessors scope their evaluation correctly.

What is covered

What is not covered

Assessor Tip

During AU-2 assessment, compare the organization's declared audit event scope against the "not covered" list above. If the authorization boundary includes Kubernetes workloads, verify whether EKS audit logs are forwarded to CloudTrail and therefore within infrastructure witnessing scope. If it includes direct SSH access (e.g., for break-glass procedures), that evidence must come from a separate source (OS-level audit logs, session recording) since it is outside the cloud API call scope. The honest scope boundaries help prevent over-reliance on a single evidence source.

11. Related Guides and References

SWT3 Guides

External References