Security, Governance, and Multitenancy for AI
Security, Governance, and Multitenancy for AI
AI multiplies data sensitivity and shared-resource risk. NCP-AI tests isolation across network, identity, storage, and Kubernetes.
Core controls
- Microsegmentation: Nutanix Flow policy, often application-centric, restricts east-west traffic between inference, training, and data services.
- Identity and RBAC: Prism Central roles, NKP RBAC, and namespace-scoped service accounts; enforce least privilege for pipelines.
- Multitenancy: separate Prism Central projects, NKP namespaces/clusters, and Objects buckets per tenant; apply quotas and network policies.
- Data protection: encryption at rest, key management, bucket policies, audit logging, and secrets management outside images.
- Governance: approved model registries, image scanning, and audit trails for data and model access.
Worked example
Two tenants share a GPU node pool. Isolate them with separate namespaces, RBAC, resource quotas, NetworkPolicy/Flow rules, and per-tenant model buckets. Do not rely on GPU partitioning alone for security; vGPU or MIG separates GPU memory but not network or storage identity. Audit access to training data and model artifacts, and log policy changes.
Traps
- Assuming a shared namespace is sufficient tenant isolation.
- Leaving inference endpoints open to flat networks.
- Storing credentials in container images or notebooks.
- Granting cluster-admin to pipeline service accounts.