Skip to content

Security

Your cluster stays yours.

We're designing Ainek for least-privilege access on customer-controlled clusters. Workloads, weights, and datasets stay in your environment. Exact permissions will be documented with design partners before any enforcement.

01 / Posture

Least privilege on a customer-controlled cluster.

Intended model: Ainek authenticates to your cluster API and metrics endpoints with scoped service identity. You own the nodes, the GPU operator, and the data path. We decide and record capacity policy — we do not become where model artifacts live.

Trust boundary

02 / Deployment & data

What should run where. What should never leave.

Topology and data categories below are the design we're building toward — so early security conversations stay concrete.

In your environment

Controllers or agents beside the cluster, scheduler integration points, and local collection of the metrics needed for observe and policy.

Ainek-operated components

Where a hybrid control plane is used: policy API, configuration, and usage state required to operate the product — never your training data plane.

May leave the cluster

  • GPU and node metrics needed for capacity decisions
  • Job and allocation metadata (identity, team, priority, request size, timing)
  • Policy configuration (quotas, priorities, fair-share rules)

Never leaves

  • Model weights
  • Dataset contents
  • Training or inference payloads

Retention for operational metrics and audit records will be agreed per design-partner deployment — only as long as needed to run and review the system.

Deploy topologies

03 / Authentication

Humans and services, separately.

Target design: admin access and cluster access as distinct identities. Roles stay coarse on purpose.

Admin authentication

Operators sign in through your identity provider when configured, or an equivalent secured admin path agreed with partners.

RBAC

Conceptual roles: Admin (policy and deployment), Team lead (team quotas and views), Viewer (read-only capacity and usage). Finer grants follow your org model.

Service identity

Cluster API and metrics access via a dedicated service account or equivalent — least privilege, rotatable, scoped to required APIs.

04 / Isolation

Policy separation is not confidential computing.

Ainek is meant to separate teams in capacity policy. Hardware and runtime isolation remain your cluster’s job.

What we do

  • Team-bound quotas, priorities, and attribution
  • Policy decisions scoped to your team identity source of truth
  • Clear ownership of who consumed which GPUs

What we do not promise

  • Confidential computing or attested enclaves — unless your stack already provides them
  • Replacing Kubernetes or Slurm tenancy primitives
  • Guaranteeing process isolation beyond what your cluster already enforces

05 / Audit

Who changed capacity policy — and when.

We're building for an audit trail from the start. Capacity governance without one is not serious infrastructure.

Policy audit log

Creates and changes to quotas, priorities, fair-share rules, and enforcement mode — with actor and timestamp.

SIEM export

Audit events designed to export to your logging or SIEM pipeline in a form your team can ingest.

Compliance certifications

We do not list certifications we have not completed. Questionnaires and architecture detail are available on request during early access.

06 / Network

Narrow APIs. Encrypted in transit.

Exact ports and CIDRs will be defined per topology with partners. The shape below is the target for a first security conversation.

Required access

Cluster API for policy integration; metrics endpoints (or push path) for GPU and scheduler signals; admin API for operators.

Egress

Hybrid designs may need controlled egress to Ainek-operated endpoints. Fully in-cluster installs keep that surface minimal or none.

Encryption

TLS in transit for control-plane and admin traffic. State we persist (configuration, usage, audit) encrypted at rest.

07 / Operations

Failure modes we design for.

Availability

Control components will be deployed for continuity appropriate to each partner environment. Exact HA topology is part of install design — not a marketing claim.

If the control plane is unreachable

Intended default: scheduling fail-open under existing Kubernetes or Slurm rules. Enforce-mode behavior documented before anyone enables it.

Incidents & contact

During private preview, contact is direct by email — not a ticket queue or enterprise support desk.

Request early access.

If security review matters for your evaluation, note it in the form. We'll cover intended topology, permissions, and data categories as part of early access.