Security

Enterprise Security FAQ

Security posture and trust boundaries for gpufirewall.io.

Do you require cluster admin access or kubeconfig?

No. GPU Firewall does not store kubeconfig and does not run kubectl or helm against customer clusters. Customers install the agent themselves (GitOps, CI, or Helm). The agent connects outbound over HTTPS only.

Do you need inbound access to our cluster network?

No. There is no inbound connectivity from GPU Firewall to customer clusters. Outbound HTTPS from the cluster to the control plane is sufficient.

What data do you collect from clusters?

Telemetry includes cluster_id, namespace inventory and labels, node inventory (optional), GPU usage signals (GPU seconds and model), and health heartbeat timestamps. No application payloads, secrets, or customer data stores are required.

How do you determine if a cluster is deleted?

Deletion is cloud confirmed: AWS, GCP, and Azure control plane events (and a reconciler fallback) set cloud_status=not_found. Telemetry staleness never implies deletion.

Can you accidentally enforce actions on the wrong namespace?

Enforcement is limited to explicitly managed namespaces (for example gfw.ai/managed=true) and explicit budgets. Unmanaged namespaces are always skipped.

What happens if pricing is unknown or delayed?

GPU Firewall does not enforce destructive actions when pricing is unknown. It records usage and warns only.

Do you support enterprise pricing and internal chargeback rates?

Yes. Finance can define rate cards per provider, region, and GPU model, including CSV import. Overrides take precedence over public list prices.

How is access controlled in the control plane?

API access uses scoped bearer tokens (admin_ui, telemetry, node_sync, cloud_events). Tokens are hashed at rest, revocable, and auditable. No session cookies are required.

Do you provide an audit trail?

Yes. Actions and configuration changes are logged with timestamps, actor identity (token), inputs, pricing source, and decision outputs.

How do you handle incident response and rollback?

Agent uninstall is customer controlled (helm uninstall). Node-sync is required for pricing and cluster actions in Kubernetes installs; webhook is optional. The control plane surfaces remediation commands for onboarding failures and provides status reports for discovery and pricing.