← Governance

Section 04 · Secure SDLC

NIST AI RMF, operationalized.

Secure AI development must be risk-proportionate, architecture-aware, adversarially informed, and operationally monitored. This framework embeds AI controls across the lifecycle while aligning to the four NIST AI RMF functions: Govern · Map · Measure · Manage. The framework scales these across three tracks driven by the risk tier.

Core stance

We do not replicate NIST AI RMF. We operationalize it inside enterprise development workflows, so governance becomes part of CI/CD instead of a parallel artifact.

Three tracks

Aligned to risk tier. Each track defines required controls under Govern, Map, Measure, and Manage. Higher tiers add controls; they don’t replace lower-tier baselines.

Baseline Track

Tier 1–2

Low-to-moderate impact systems. Lightweight controls, recommended adversarial testing.

Govern

  • Risk tier documented
  • Ownership assigned
  • Data classification confirmed

Map

  • Architecture diagram created
  • Data sources documented
  • Third-party dependencies identified

Measure

  • Basic performance metrics defined
  • Logging enabled
  • Monitoring thresholds defined

Manage

  • Incident reporting path documented
  • Model versioning enforced

Enhanced Track

Tier 3

Customer-facing, sensitive, or regulatory-exposed systems. Adversarial integration begins here.

Govern

  • Governance validation required
  • Vendor AI review formalized
  • Bias evaluation summary documented

Map

  • Formal threat model completed
  • ATT&CK / ATLAS mapping for relevant adversarial techniques
  • Automation flow documented

Measure

  • Pre-deployment evaluation suite (capability, safety, regression baselines, jailbreak)
  • Drift detection thresholds defined
  • Abuse case simulations performed
  • Logging validated for forensic use

Manage

  • Escalation triggers defined
  • Model rollback procedure tested
  • Red team testing performed where applicable

High-Assurance Track

Tier 4

Regulated decisions, financial or employment impact, fully automated systems, or high-reputation risk.

Govern

  • Executive visibility
  • Formal risk acceptance documentation
  • AI risk committee review (where maturity allows)

Map

  • Full architecture threat modeling
  • ATLAS adversarial technique mapping
  • Third-party supply chain risk mapping
  • Documented training data provenance and lineage (sources, licensing, consent basis, contamination checks)

Measure

  • Structured adversarial testing
  • Bias and fairness analysis
  • Explainability documentation
  • Continuous drift monitoring dashboard
  • Pre-deployment evaluation suite required, with regression gating on release

Manage

  • Real-time anomaly alerting
  • Incident response integration
  • Regulatory reporting playbook
  • Periodic model review cadence

Control inheritance

Don’t make product teams re-implement platform controls.

Controls may be inherited from platform-level logging enforcement, the centralized model registry, infrastructure security baselines, and vendor due diligence templates. Product teams only implement controls not already inherited. This is the difference between governance that scales and governance that creates duplicate work.

Tier 3+ requirement

Vendor AI due diligence checklist

For any third-party model, fine-tuning service, or AI-powered vendor capability. Failure on any item is not automatic disqualification; it is a documented risk acceptance.

01Model card review (capabilities, limitations, evals)
02Training data disclosure and licensing posture
03Customer data isolation (no cross-tenant training)
04Fine-tuning data handling and retention
05Eval transparency (published benchmarks and methodology)
06Change-notification SLA for model updates
07Version pinning and rollback support
08Indemnification and liability terms for AI outputs
09Subprocessor disclosure and cross-border processing
10Security certifications (SOC 2 Type II, ISO 27001)

Executive reporting (Tier 3–4)

  • Active AI systems by tier
  • Drift incidents
  • AI-related security incidents
  • Vendor AI dependencies
  • Open risk acceptances
  • Pre-deployment eval pass rates

Risk visibility, without micromanaging product teams.