CVE-2026-68903: LLMKube accepted any absolute path or file:// URI as a model source, letting a namespace user copy node files or the controller's service-account token into model storage (fixed 0.9.0)
LLMKube is a Kubernetes operator for self-hosted LLM inference (llama.cpp, vLLM, TGI, mlx-server). Before 0.9.0, its ModelReconciler accepted arbitrary absolute paths and file:// URIs in Model.spec.source with no allowlist. A namespace user who can create or update Model and InferenceService resources could make copyLocalModel open any file in the controller pod (such as the controller's own ServiceAccount token) and copy it into the model-cache PVC they can read, or, with model caching enabled (the default in the CLI and Helm chart), get a HostPathVolumeSource that mounts any node file (such as /etc/kubernetes/kubelet.conf) into the inference pod. That exposes credentials outside the attacker's namespace. Fixed in 0.9.0.
What it is
Entry point: the kubebuilder validation pattern on Source in api/v1alpha1/model_types.go explicitly allows absolute paths; isLocalSource/getLocalPath in internal/controller/source.go only strip the file:// prefix, with no canonicalisation.
Sink 1: os.Open(localPath) in copyLocalModel (internal/controller/model_controller.go) runs in the controller pod. Sink 2: a hostPath volume with the attacker's path mounted into the inference pod's init container.
Affected: github.com/defilantech/LLMKube <= 0.8.28; fixed in 0.9.0 (PR #981). Reported by EQSTLab and knm6777.
GitHub published the advisory on 8 July 2026; NVD published the CVE on 18 November 2026.
How to check
- Check the LLMKube operator version (Helm release or controller image tag).
- List Model resources whose spec.source starts with / or file:// and confirm each path is one you expect.
- Review who can create or update Model and InferenceService resources in each namespace.
How to fix
- Upgrade LLMKube to 0.9.0 or later.
- Until then, restrict create/update on Model and InferenceService to trusted users, or turn off modelCache.
- If untrusted namespace users had access on an older version, rotate the controller's ServiceAccount token and any node credentials that could have been mounted.
kubectl get models.inference.llmkube.dev -A -o jsonpath='{range .items[*]}{.metadata.namespace}/{.metadata.name}: {.spec.source}{"\n"}{end}' 2>/dev/null \
|| kubectl get models -A -o custom-columns=NS:.metadata.namespace,NAME:.metadata.name,SOURCE:.spec.source