Which frameworks should guide AI data security and model governance?

Why This Matters for Security Teams

AI data security and model governance fail when they are treated as separate problems. The data pipeline, training artifacts, deployed models, and the identities that access them are one control surface, so gaps in any layer can expose sensitive data or degrade model trust. NIST’s Cybersecurity Framework 2.0 gives teams a common risk vocabulary, while NIST AI RMF helps frame model risk and governance decisions.

That matters because non-human identities often become the hidden path into AI systems. NHIMG’s Ultimate Guide to NHIs — Key Research and Survey Results notes that 79% of organisations have experienced secrets leaks and 96% store secrets outside secrets managers in vulnerable locations. In practice, many security teams discover model compromise only after a secret, dataset, or service account has already been abused rather than through intentional governance design.

How It Works in Practice

The strongest approach is to map the AI lifecycle to governance, assurance, and access controls instead of trying to secure “the model” in isolation. NIST AI RMF is useful for defining risk treatment across data sourcing, training, testing, deployment, monitoring, and incident response, while the NIST Cybersecurity Framework 2.0 helps align those activities with enterprise security outcomes such as asset management, access control, detection, and recovery.

For identity and access, OWASP NHI guidance is important because AI workloads depend on service accounts, API keys, tokens, and automation identities. NHIMG’s Ultimate Guide to NHIs emphasises that NHIs outnumber human identities by 25x to 50x in modern enterprises, which means model governance can fail at scale if these identities are not inventoried and controlled.

• Classify training data, prompts, embeddings, and model outputs by sensitivity and retention need.
• Bind model and pipeline actions to least-privilege service identities, not shared admin accounts.
• Require secrets rotation, short-lived credentials, and revocation for CI/CD, MLOps, and inference services.
• Log model access, data access, and administrative changes in a way that supports audit and incident response.
• Validate model approvals, dataset provenance, and change management through policy-as-code where possible.

Current guidance suggests this works best when governance owns the model risk decisions, security owns the identity and telemetry controls, and engineering owns the pipeline enforcement points. These controls tend to break down when AI systems are embedded across many business units because identity sprawl and inconsistent logging make end-to-end accountability difficult.

Common Variations and Edge Cases

Tighter governance often increases engineering friction, requiring organisations to balance faster experimentation against stronger control over data, prompts, and release paths. That tradeoff is real in research environments, but it should not justify permanent exceptions for production systems.

Best practice is evolving for frontier models, multi-agent pipelines, and external model hosting, so there is no universal standard for this yet. In those cases, teams should lean on the Ultimate Guide to NHIs — Standards to anchor identity and lifecycle controls, then adapt NIST AI RMF governance to the specific deployment model. The NIST AI RMF is especially useful where model behaviour changes after retraining or where vendors retain operational control.

Edge cases also arise when regulated data, third-party APIs, and embedded copilots all touch the same workflow. In those environments, security teams should not assume a single control framework will cover everything. Instead, combine CSF 2.0 for enterprise security outcomes, AI RMF for model governance, and NHI controls for machine identities. In practice, many organisations only notice the gap after a leaked token, an over-permissioned pipeline, or a model update has already widened access.

Related resources from NHI Mgmt Group

• Why is single-provider AI agent governance not enough for enterprise security?
• Why do AI systems create governance gaps that standard app security misses?
• Which frameworks should guide identity governance for human and non-human identities?
• What do security and AI teams get wrong about model collapse?