Admission Control and Image Integrity
Hardening a chart is only useful if the settings survive contact with a real cluster. OLIVE therefore supports admission-time policy enforcement, and ships the supply-chain evidence that makes it meaningful.
The chain, end to end
- Dependencies are pinned. All npm packages are pinned to exact versions with lockfile integrity verification, strict resolution, and rejection of unexpected content changes for a resolved version. A seven-day cooldown prevents the build from adopting a third-party version published in the last week — the cheapest effective control against the dominant npm attack pattern, where a compromised maintainer account publishes a malicious version that is discovered and unpublished within hours to days.
- Images are pinned by digest. Every container image, first- and third-party, is referenced by its SHA-256 manifest digest rather than by a tag. A tag is a mutable pointer: anyone who can write to the registry can repoint
v1.2.0at different content, and every cluster that pulls it runs something nobody reviewed, with no change in the manifest and nothing in the deployment history to show it. - Images are signed. OLIVE signs the images it builds with cosign (Sigstore). Digest pinning proves the image running is the image the manifest named; it does not prove the image the manifest named is the image OLIVE built. Signing binds the artefact to the identity that produced it.
Signatures are keyless: there is no stored public key, so the verifier checks the signing certificate against Fulcio and the transparency-log entry against Rekor (https://rekor.sigstore.dev by default). The Kyverno pods therefore need HTTPS egress to fulcio.sigstore.dev, rekor.sigstore.dev and tuf-repo-cdn.sigstore.dev. Both calls happen inside the admission webhook, synchronously, while a pod is being created — which is why the policy raises webhookConfiguration.timeoutSeconds to 30.
The consequences for your cluster:
- The Kyverno pods need egress to Fulcio and Rekor over HTTPS. If general egress is restricted, carve it out for those pods specifically.
- On a cluster with no route to them,
enforcewithfailurePolicy: Failblocks every pod creation in an enrolled namespace. WithIgnoreit fails open and checks nothing. - A private Sigstore deployment works: point
ctlog.urlat your own Rekor, and change the keyless identity to your own signing pipeline.
Signature verification is a switchable control and off by default, so this is a prerequisite for enabling it rather than for running OLIVE. The registry allow-list makes no network call and is unaffected.
Because enforcement means failurePolicy: Fail, an outage or a blocked route at Fulcio or Rekor stops pod creation in every enrolled namespace — so OLIVE's recommended posture is the registry allow-list at enforce and signature verification at audit, with the report feeding your alerting rather than the webhook blocking admission. If a tenant does run verification at enforce, run at least two Kyverno replicas and monitor the egress. See Admission Policies → The recommended posture and Network requirements.
- The cluster verifies. A Kyverno policy verifies the signature at admission and refuses anything else.
- An SBOM exists per release. Service-level SBOMs from dependency analysis, image-level SBOMs in SPDX format, and third-party images covered separately — so that "is any tenant running an affected version, and which?" is answerable in minutes rather than days.
Helm chart dependencies cannot currently be pinned by OCI digest. Pinning images alone does not fully protect against a modified chart: if a chart tag is mutable, the chart could be replaced and reference different images. The residual risk is reduced because the charts live in an OCI registry under OLIVE's control. Mirroring third-party images and charts into a registry of your own would remove the dependency on external registries, and some organisations choose to; it is not a documented or supported OLIVE procedure.
Installing Kyverno
Kyverno is installed as cluster infrastructure, with the engine, its CRDs and the webhook configuration. OLIVE offers the kyverno-infra chart for it, which also renders the policies described below.
The install itself — the chart values, the registry pull credential Kyverno needs in its own namespace, and the Argo CD adaptations that keep Argo and Kyverno from fighting each other — is documented in Infrastructure Charts → Kyverno, so there is one copy of it.
The policies
Image signature verification
Three modes, selected per namespace by the platform administrator through tenant-setup.imageVerification.mode:
| Mode | Behaviour |
|---|---|
disabled | The namespace is not labelled and not evaluated at all |
audit | Signed images are verified; images that fail verification are admitted but recorded |
enforce | Images that cannot be verified as properly signed are not deployed |
Registry allow-list
A separate policy and a separate label, answering a different question: not did our pipeline sign this but is this image from a registry we approved at all. Complying with the allow-list needs no rebuild, so tenants can usually reach enforce here long before signatures are ready.
Running both is the defence-in-depth position: a validly signed image from an unexpected source is still refused.
cluster-operator-written policies
Policies of your own can be applied selectively using namespace labels, so a policy can target specific tenant namespaces rather than the whole cluster. Any organisation with its own policy baseline can apply it to OLIVE namespaces without a change to the OLIVE charts, and without being forced to apply it uniformly across tenants with different requirements.
Reading policy reports
# All results in a namespace
kubectl get policyreport -n oli-single
# Detail of one report
kubectl describe policyreport <policy-report-id> -n oli-single
A failing entry looks like:
Results:
Category: Supply Chain Security
Message: Image is not signed by the Olive signing pipeline (...)
Policy: require-signed-olive-images-audit
Result: fail
Severity: high
Use the reports to see exactly what would be blocked before switching a namespace to enforce.
enforce fails lateAdmission control acts on pod creation. Switching to enforce evicts nothing and everything looks fine — until the next rollout, scale-up, eviction or node drain. A tenant can sit in a broken state for weeks and discover it during a node reboot at 2 a.m. A crash-looping container does not recreate the pod object and is therefore not re-admitted, which is why the failure is so delayed.
Move a namespace to enforce only after its audit report has been empty across a full deployment cycle.
Deployment control
Deployment runs through Argo CD on a GitOps model: the desired state lives in a repository and Argo CD reconciles the cluster against it. The security-relevant parts are which charts and versions deploy, registry access, secret references without plaintext in Git, synchronisation and approval behaviour, rollback, drift detection, deployment permissions, and auditability of who changed the desired state.
Automated deployment must not mean that every repository change reaches production unreviewed. Agree the promotion and approval workflow explicitly, and keep drift detection on: a control that is enforced at deploy time but silently drifts afterwards is not a control.