From Interview Questions to Cluster Damage: Adventures in k8s Cluster Shenanigans

The talk grew out of the questions the presenters ask job candidates in security interviews: a free-form game about how someone would disrupt or quietly misuse a working Kubernetes cluster. The speakers opened with a short primer on core objects such as pods, deployments, namespaces, service accounts and secrets. They pointed out that secrets are only base64 encoded and that namespaces are a weak boundary that depends entirely on role-based access control.

The middle of the session walked through several scenarios. Taking down CoreDNS breaks new name lookups across the cluster, yet image pulls still worked because the kubelet resolves names through the node's own resolver. Pinning CoreDNS to a node label and scaling its replica set to zero proved more stubborn than expected, since a leftover pod kept DNS alive until both were changed. A malicious validating admission webhook was then shown forwarding every new secret or pod definition to an external server. Finally, a monitoring tool's Prometheus endpoint exposed cluster-wide inventory in a single request, which is hard to spot in normal traffic.

The takeaways were practical: restrict who can change kube-system components and webhooks, alert on webhook configuration changes, use network policies to limit workload access, and verify that security products actually monitor what their vendors claim.