bedrock assumes that any single control can fail. Security comes from several independent layers, so that getting past one still leaves the others in place.
#Supply chain
Every upstream bundle (Longhorn, Kyverno, CloudNativePG, prometheus-operator, Strimzi) and the RKE2 release itself is downloaded with a sha256 checksum and refused if it does not match. Every container image in the repository is pinned by digest, so the content cannot change behind a tag. A Kyverno policy allows images only from a short list of publishers.
#Admission
- Pod Security Admission enforces the
restrictedlevel on almost every namespace. Kubernetes itself refuses root, privileged and host-path pods there, without depending on Kyverno. - Kyverno requires pinned tags, CPU and memory limits and allowed publishers on application pods, and fills in non-root defaults.
- A ValidatingAdmissionPolicy runs inside the API server and keeps the high priority classes for storage, databases and system components. It works even when Kyverno is down.
#Network
Default deny everywhere, flows pinned on both sides and no exposed ports. See networking.
#Identity and access
No component has wildcard RBAC. Default service accounts do not mount API tokens. kube-state-metrics, Grafana’s sidecar and Reloader have no access to Secrets. The database replicas read the primary’s backups with a read-only key, so a compromised replica cannot delete them. PostgreSQL’s superuser cannot log in over the network, and every database client must use TLS.
#Runtime
Falco watches system calls on every node: a shell opening in a database pod, an unexpected binary, an attempt to break out of a container. Events go to Loki and to Prometheus, and critical ones raise an alert.
#Data at rest
Kubernetes Secrets are encrypted in etcd. Every Longhorn volume is encrypted with LUKS: the databases, the monitoring stack, the event buses, Garage, which holds every backup, log and trace, and OpenBao.
#Threat model
The table shows what an attacker who gains a shell in one pod can reach.
| Foothold | Can reach | Cannot reach |
|---|---|---|
| An application pod in its own namespace | Only the services its own policy names, such as its database pooler | Other tenants, the monitoring stack, Garage, other databases |
| Grafana | Prometheus, Loki, Tempo and the Kubernetes API | Alertmanager, Garage, databases |
Alloy in node-agents | Loki on 3100, Tempo on 4317 and the Kubernetes API | Alertmanager, Garage, databases |
node-exporter in node-agents | It runs on the host network, so Cilium treats it as the node itself: whatever each namespace lets the node reach, such as kubelet probe ports | Ports a namespace does not open to the node |
| A database replica | Its own cluster, Garage on 3900, the primary on 5432, the Kubernetes API | The primary’s backups for writing, other tenants |
A breach in one tenant does not reach another, because tenants share no cluster, node or network.
#Accepted exceptions
Some findings are left in place on purpose. Each has a reason.
| Finding | Why it stays |
|---|---|
| CIS 5.1.6: some pods mount API tokens | They are operators and controllers that need the Kubernetes API to work. The exception is documented, not hidden by editing the benchmark. |
| Kyverno can write Secrets cluster-wide | It copies the replication certificates into the replica namespaces and keeps them current when they rotate. Its ConfigMap writes are limited to bedrock-failover by a Role; cluster-wide it can only read ConfigMaps. |
| Pod traffic on the same node is not encrypted | It never leaves the machine. |
| A placeholder Longhorn encryption key in the repository | It exists so a fresh deploy works. Choose its value before the first deploy since it cannot be rotated on a running tenant. Every other platform secret is made at bootstrap and kept in OpenBao or made by its own operator. |
| Garage accepts secret files others can read | The CSI driver writes files with mode 0644, so GARAGE_ALLOW_WORLD_READABLE_SECRETS is set. The files sit in memory inside the pod and Garage is the only process there. |
#Compliance
The cluster is checked with kube-bench against the RKE2 CIS 1.9 benchmark. Every check passes except 5.1.6, the exception above. The full self-assessment is in the repository at reports/cis-1.9.pdf.