Go · github.com/open-feature/open-feature-operator
open-feature-operator: Cross-namespace FeatureFlagSource and InProcessConfiguration resolution exposes spec contents on multi-tenant clusters
A namespaced FeatureFlagSource or InProcessConfiguration resource can be referenced cross-namespace via the openfeature.dev/featureflagsource annotation using the documented {NAMESPACE}/{NAME} syntax. The operator resolves the referenced resource cluster-wide and materializes its contents (env vars, flagd sidecar arguments including httpSyncBearerToken, sync URIs, supporting ConfigMaps) into the referencing workload.
On multi-tenant clusters that treat namespaces as trust boundaries, a tenant who can deploy a controller-owned workload in their own namespace can cause the operator to read another tenant's FeatureFlagSource / InProcessConfiguration spec contents.
FeatureFlagSource / InProcessConfiguration spec, including spec.envVars literal values, spec.httpSyncBearerToken, and sync URIs.The cross-namespace {NAMESPACE}/{NAME} annotation syntax is intentional and documented in docs/annotations.md and docs/feature_flag_source.md. The operator's cluster-wide RBAC scope is intentional. Namespace-as-trust-boundary is not part of the operator's current stated security model.
This advisory makes the tenancy assumption explicit and tracks the architectural change that will eliminate the implicit cross-namespace pattern.
Two technical points in the original report require correction:
secretKeyRef / configMapKeyRef cross-namespace disclosure is not possible via this path. Kubelet resolves these as LocalObjectReference against the pod's own namespace; the operator does not bypass that. The actual disclosure surface is FeatureFlagSource / InProcessConfiguration spec contents the operator itself materializes (inline envVars values, httpSyncBearerToken, sync URIs).create featureflagsources is not a prerequisite. The webhook rejects pods without OwnerReferences (pod_webhook.go:75-77), so the prerequisite is create on a workload controller (deployments, statefulsets, daemonsets, jobs, cronjobs, replicasets) in a namespace the attacker controls. FeatureFlagSource create in any namespace is not required.As with any Kubernetes CRD, treat the spec content of FeatureFlagSource and InProcessConfiguration as readable by anyone with read access to the resource, and don't place plaintext secrets in CR spec fields. Fields most likely to bite users:
spec.sources[].source, when the URI embeds credentials (e.g. https://user:pass@host/repo)spec.sources[].certPath, if the path itself is sensitivespec.envVars[].value (use valueFrom.secretKeyRef instead; kubelet enforces same-namespace resolution and the secret value is not stored in the CR)If developers treat namespaces as trust boundaries:
create on featureflagsources / inprocessconfigurations via RBAC where feasible,A future release will introduce explicit cluster-scoped CRDs (ClusterFeatureFlagSource, ClusterInProcessConfiguration) and remove implicit cross-namespace resolution. This is a breaking change tracked in #847.
This class of issue (authenticated namespace tenant abuses an unenforced cluster-wide surface that crosses an assumed namespace boundary) has Kubernetes precedent: CVE-2020-8554 (External IPs) was accepted as documented posture and mitigated via an opt-in admission plugin.
Reported by @0xVijay. Thanks for the disclosure. This appears to be an example of https://cwe.mitre.org/data/definitions/668.html. In terms of how it ended up here, it's more of an unimplemented security feature than an "bug". It seems to deviate from reasonable expectations and conventions in the K8s ecosystem. See https://nvd.nist.gov/vuln/detail/cve-2020-8554 as an example of a comparable vulnerability.
Is your project exposed to this? Stateward checks every dependency on every pull request and flags it only if your code actually reaches it.
Check my repoSources: CISA KEV (public domain), OSV.dev & GitHub Advisory Database (CC-BY-4.0), FIRST EPSS, NVD/CWE (public domain). Served live from the Stateward advisory database.