The NHI Governance Framework
Seven layers. One question each.
Governing an AI agent means answering seven questions, continuously: who it is, what it can prove, what it may do, whether it is still what you approved, what it did, whether it should still exist, and who answers for it. The framework names each layer so teams can assess where they stand.
Identity and registry
Who is this agent?
A registered identity bound to its purpose, risk tier, allowed models and a named owner. Unregistered callers are discovered, not trusted.
Credentials and delegation
What can it prove right now?
Short-lived signed credentials, keys in an HSM or KMS, delegation that only narrows, and fast revocation.
Authorization
May it make this call?
Per-call decisions over tools, arguments, data scopes, rates and budgets, with irreversible actions held for a person.
Runtime attestation
Is it still what we approved?
Model and workload attestation, behavioral baselines, and content-security signals that narrow an agent that drifts.
Audit and forensics
Can we prove what happened?
An append-only, hash-chained record written before actions run, with signed checkpoints and trusted timestamps.
Lifecycle
Does it still need to exist?
Recertification, orphan and dormancy detection, reviewed capability changes and clean retirement.
Human oversight
Who is accountable?
Rules for when a person approves, edits or is notified, plus a kill switch that stops any agent in one step.
One decision record
Each decision records the result and reason of every layer, so engineers, approvers and auditors read the same story.
Self-assessment
Where does your agent program stand?
For each layer, ask whether you can answer the question for every agent, today, with evidence.
- Can you list every agent, its owner and its permissions?
- Do agent credentials expire within minutes, with keys in an HSM?
- Is every tool call checked before it runs, including arguments?
- Would you notice an agent behaving differently from last week?
- Could you show an auditor exactly why an action was allowed?
- Are agents without an owner, or unused for months, retired?
- Can a person stop any agent in one step, and have you tested it?
Mapped to standards
Evidence, not a checkbox
Sentinel maps each layer to controls in SOC 2, ISO/IEC 42001, the NIST AI Risk Management Framework and the EU AI Act, and produces a signed evidence bundle for your assessor.
Evidence mappings support an assessment. They are not a certification or a legal determination of conformity.
Security and trustWalk through the seven layers with us.
A one-hour workshop: we assess your agent program against the framework and show how Sentinel closes the gaps.