Skip to navigation

Governance and Assurance

View as Markdown
View cross-stack controls in the interactive architecture

Governance and assurance are cross-stack architecture responsibilities. Controls must exist at enforceable identity, network, runtime, data, artifact, and administrative boundaries rather than relying on model behavior alone. Architecture owners define risk, approved policy, accountable authority, and acceptable evidence. Platform and operations owners implement, monitor, maintain, and recover the controls that satisfy them.

This layer defines what must be governed and assured across the architecture: risk, approved policy, compliance, control objectives, accepted evidence, and accountable identity and delegation decisions. Platform Services implement identity and trust, policy enforcement, supply-chain controls, observability, detection, and audit.

Explore security and governance ecosystem solutions.

Control Planes

Resolve the following decisions across each governance and assurance plane:

PlaneRequired Decisions
Identity and delegation accountabilityDistinguish users, services, workloads, workspaces, agents, and tool calls. Bind actions to accountable sponsors, approved delegation, autonomy limits, human-review gates, and acceptance evidence. Delegation must preserve the parent authority boundary. Refer to SAW identity and access boundary.
Policy assuranceDefine required policy properties, verification coverage, assumptions, exceptions, and acceptance evidence. Platform Services checks policies; runtime owners establish that enforcement matches approved policy. Refer to OpenShell policy verification.
NetworkDefine ingress, egress, east-west paths, isolation, inspection, service exposure, default-deny policy, and restricted destinations. Refer to OpenShell policies.
Credentials and secretsBroker bounded, short-lived credentials. Define storage, conditional release, rotation, revocation, and audit. Refer to OpenShell credential providers.
Data governance and protectionAssign ownership and classify sources and derived artifacts. Define approved use, privacy, residency, retention, access, lineage, persistent memory and context protection, and disclosure requirements. Protect inference inputs, retrieved context, and outputs according to their sensitivity. Refer to Data and action boundaries.
Models and artifactsEstablish provenance, scanning, signing, immutable identity, promotion, withdrawal, model intellectual-property protection, and supply-chain controls. Refer to NVIDIA Secure AI Operations Guide.
RuntimeSelect an approved container sandbox, microVM, VM, or equivalent isolation boundary. Constrain processes, files, tools, devices, network access, and privileged operations below the model trust boundary. Refer to OpenShell runtime architecture.
Platform assuranceDefine measured state, attestation appraisal, conditional key release, approved exceptions, evidence freshness, and failure behavior where assurance depends on platform state. Refer to NVIDIA Attestation.
Monitoring and responseCorrelate identity, policy decisions, tool actions, data access, model requests, runtime activity, deployments, vulnerabilities, and infrastructure events. Define containment and recovery ownership. Refer to AI Factory Observability Guide.

Enforcement Boundaries

Architecture-wide requirements must resolve to enforceable controls and named operators:

  • Platform Services provide Identity and Trust Services, Policy and Enforcement, Delivery and Supply Chain, and Observability and Detection for approved requirements.
  • Cloud-Native Platform provides admission, network policy, controlled egress, workload identity, tenant isolation, sandbox or virtualization boundaries, runtime telemetry, patching, and containment primitives.
  • Operations and Resilience owns secure maintenance, vulnerability remediation, incident response, continuity, recovery, and the day-two lifecycle of every enforcement point.
  • Capability Reference Designs compose those controls for a defined problem. Secure Agent Workspace, for example, requires an approved agent-isolation boundary, workspace identity, brokered credentials, network perimeter controls, and attributable audit.

Do not assign a named product to the public architecture until its role, boundary, dependencies, support model, and accountable owner are approved. This keeps the core stable while product-specific implementation guidance evolves.

NVIDIA Open Agent Safety Platform is a related NVIDIA reference design for governing and monitoring autonomous agents across software and hardware boundaries. It combines NVIDIA OpenShell runtime controls with NVIDIA Sentry and supported NVIDIA infrastructure. It is linked guidance rather than a core element or Capability Reference Design in this publication. Use its technical architecture overview to assess applicable runtime-governance, out-of-band monitoring, and infrastructure-enforcement patterns.

Model and Inference Confidentiality

Define who can access model weights and disclose inference inputs, retrieved context, and outputs. Apply those decisions to model storage, delivery, serving, telemetry, and retirement. AI Platform owns the serving implementation; Platform Services mediates access and disclosure. Governance and Assurance defines protection objectives and accepted evidence.

Confidential computing can protect selected material in use when supported by the deployment and threat model. It complements access, disclosure, and lifecycle controls rather than replacing them.

Confidential Computing

Use confidential computing when the threat model requires stronger isolation or verifiable protection for selected data, models, credentials, or workloads while they are in use. It introduces platform-specific hardware, firmware, driver, attestation, key-release, monitoring, and recovery requirements. It is not a blanket security upgrade for every workload.

Use the following deployment boundaries to assess architecture requirements:

Deployment BoundaryArchitecture Consideration
Confidential virtual machineProtect the guest boundary and verify the measured VM, host, and accelerated-compute state before releasing sensitive material. Refer to confidential inference on self-hosted virtual machines.
Confidential containerCombine container and workload identity with a supported confidential-computing boundary. Do not treat an ordinary container namespace as equivalent isolation. Refer to confidential inference on self-hosted Kubernetes.
Bare-metal confidential computingUse only where current platform guidance explicitly supports the target hardware and release. Verify ownership of firmware, attestation, key release, and recovery. Refer to NVIDIA Trusted Computing Solutions.

Begin with the NVIDIA Trusted Computing Solutions library for current deployment options, attestation, compatibility, and operational guidance. Do not rely on roadmap status or an unverified platform combination in an architecture approval.

1

Measure

Collect evidence describing firmware, boot state, host or guest boundary, and GPU state.
2

Verify

Evaluate evidence against explicit policy before treating the environment as trusted.
3

Release

Release keys, encrypted models, or sensitive data only to an approved measured environment.
4

Operate

Monitor trust state, lifecycle changes, revocation, incidents, and recovery throughout execution.

Governance Outcomes

  • Named owners for architecture, platform, security, data, application, and operations decisions.
  • Approved exceptions with scope, risk, compensating controls, and review date.
  • Traceable changes to policies, models, prompts, tools, connectors, and deployment state.
  • Human-review gates proportional to workload authority and potential impact.
  • Defined vulnerability-management, patch, containment, and recovery responsibilities for platform and workload owners.
  • A clear statement of which organization owns compliance authorization and product support.