Skip to main content

Security Configuration

OLIVE 1.3 ships a set of security controls that fall into two groups, and the distinction matters more than any individual setting:

  • Always on. Controls built into the services and the charts. You cannot switch them off from values.yaml, and you do not have to think about them: independent JWT validation in every service, the egress guard that blocks calls to internal and metadata addresses, per-user authorization on jobs, sessions and files, input validation and limits, path-traversal prevention, container hardening (non-root, read-only root filesystem, dropped capabilities, seccomp, AppArmor, no privilege escalation), digest-pinned images and pinned dependencies. The only exceptions are global.appArmor and global.seccomp, which exist because some hosts cannot enforce those two profiles — see Workload Hardening.
  • Switchable by you. Controls that depend on what the cluster provides, or that have an operational cost: network policies, service mesh mTLS, admission policies, the service-account model, credential rotation, and Keycloak's authentication policy.

This section covers the second group — the controls a platform administrator decides on and configures. What the cluster has to provide for them to work is described in the Cluster Operator section.

The platform makes it possible; the administrator makes it real

Almost every switchable control defaults to off or to a reporting-only mode, because a control that silently breaks a deployment is worse than one you enable knowingly. In a regulated environment they are expected to be on. Nothing in this section requires a code change or a custom build — it is all deployment values and Keycloak realm configuration.

What you configure​

PageControls
Network Policiesglobal.networkPolicy.*, global.keycloakIpBlocks, global.nonRoutableCidrs
Service Mesh and mTLSglobal.serviceMesh.*
Admission Policiestenant-setup.imageVerification.mode, tenant-setup.imageRegistry.mode, Pod Security Admission
Identity and Secretsglobal.serviceAccounts.mode, Vault roles, credentials.rotationId, Keycloak realm policy, OLIVE roles
Security LoggingLog streams, retention, and who may read the security stream

Set the values you want, then deploy​

Security controls are part of the deployment values, so the normal path is to decide the posture first and deploy with it in place — not to deploy and then switch controls on one by one. An installation of 1.3 is a fresh deployment (see Migration 1.2.0 to 1.3.0), so there is no staged migration to work through.

Start both admission modes at audit, since a control in a reporting mode costs nothing and tells you a great deal. Where they should end up differs, and that difference is a recommendation rather than a staging order:

  • tenant-setup.imageRegistry.mode and tenant-setup.podPrivileges.mode → enforce. Neither makes a network call, and complying with them needs no rebuild.
  • tenant-setup.imageVerification.mode → audit, as a defensible permanent position. Enforcing it requires failurePolicy: Fail, which puts a synchronous dependency on Fulcio and Rekor — two public internet services — into the pod-creation path. See Admission Policies → The recommended posture.
  • global.serviceMesh.mode at none unless a mesh is actually installed on the cluster and you have confirmed which mode it runs.

If something does not work, narrow it down in this order​

These controls interact, and their failure modes are mostly silent — a timeout rather than an error, a policy that evaluates nothing, a probe that never passes. When a deployment misbehaves and you suspect the security configuration, work through them in this sequence. It is ordered by how likely each one is to be the cause and how cheap it is to rule out, so the first steps are quick and the expensive one comes last:

  1. Pod Security Admission — check for violations first: kubectl get events -n <ns> --field-selector reason=FailedCreate. A rejected pod is an obvious cause with an obvious message.
  2. Admission policies — if pods are refused at creation, drop imageVerification.mode to audit and re-check. kubectl get policyreport -n <ns> shows what a policy would have blocked.
  3. Service accounts — if a workload cannot reach Vault, confirm it is presenting the identity its Vault role names. In split mode each role accepts exactly one ServiceAccount.
  4. Network policies — the most common cause of "everything times out". Set global.networkPolicy.enabled: false temporarily: if the symptom disappears, a rule is missing. Network Policies lists the usual suspects.
  5. Service mesh — rule this out only after network policies, because the two interact. A serviceMesh.mode that disagrees with the cluster produces policies for the wrong traffic shape, and probe failures right after enrolling a namespace are almost always the ambient health-probe rule.

The same sequence works as a bring-up order on an unfamiliar cluster, if you would rather introduce the controls a few at a time than diagnose them later. That is a choice about how much risk you want in one step, not a required procedure.

Order matters between network policies and the mesh

serviceMesh.mode and networkPolicy.enabled are not independent. Ambient mode tunnels traffic over HBONE on TCP 15008, so every policy whose peer is a meshed pod must also allow that port; sidecar mode needs the opposite change. The charts derive this from serviceMesh.mode — which means the value has to match what is actually installed on the cluster. If it does not, you get timeouts rather than errors.

Keep the security values in one place​

networkPolicy, serviceMesh, nonRoutableCidrs, keycloakIpBlocks and vault are cluster-wide facts, not per-tenant preferences. Put them in a single shared values snippet that every deployment includes. A tenant whose serviceMesh.mode disagrees with the shared runtime pool will render policies for the wrong traffic shape.

What this section does not cover​

  • Cluster hardening — node configuration, CNI choice, cloud IAM, perimeter firewalls, encryption at rest. That is the cluster operator's, described under Cluster Operator → Security.
  • Detection and response — OLIVE emits the security events; alert rules, triage and incident response are operational capabilities you build on top. See Log Events and Automated Reactions.
  • Backup and disaster recovery — see What to Back Up.
  • Troubleshooting — the same failure modes collected by symptom, together with the ones outside the security configuration