A high-severity vulnerability in LLMKube, a Kubernetes operator for self-hosted large language model inference, allowed any namespace user with permission to create or update Model resources to read arbitrary files from the controller pod or the underlying node.
Quick Take
- What: CVE-2026-68903, a path-traversal vulnerability in LLMKube versions before 0.9.0 that accepts arbitrary absolute paths and file:// URIs in Model.spec.source without restricting them to approved roots.
- Who is affected: Kubernetes clusters running LLMKube 0.8.28 or earlier with model caching enabled (the default configuration).
- CVSS score: 7.7 (High).
- Fixed version: 0.9.0, released July 5, 2026.
- What to do: Upgrade to LLMKube 0.9.0 or later.
In plain English
LLMKube is a tool that helps run AI language models on Kubernetes clusters. Before version 0.9.0, it had a security hole: anyone allowed to create or edit model definitions in their namespace could trick the system into reading sensitive files, such as the controller’s own security credentials or files from the server itself. The attacker could then copy those files to storage they control and read secrets from other parts of the cluster. The developers fixed this in version 0.9.0 by blocking unapproved file paths.
What you should do: If you run LLMKube, you should upgrade to 0.9.0. If you do not use LLMKube, you can ignore this issue.
What happened
The ModelReconciler component in LLMKube accepted arbitrary absolute paths and file:// URIs in the Model.spec.source field without validating them against an allowlist. A namespace user who could create or update Model and InferenceService custom resources could exploit two distinct code paths.
The first sink, copyLocalModel, called os.Open() on the attacker-supplied path inside the controller pod itself. An attacker could set model.spec.source to /var/run/secrets/kubernetes.io/serviceaccount/token, causing the controller to open its own Kubernetes service-account token and write it to the model-cache persistent volume claim, which the attacker could then read.
The second sink created a Kubernetes HostPathVolumeSource that mounted the attacker-selected path from the node filesystem into the inference pod’s init container. When model caching was enabled (the default in both the CLI and the official Helm chart), the init container copied the node file to the model-cache PVC. An attacker could read sensitive node files such as /etc/kubernetes/kubelet.conf.
The vulnerability was present in LLMKube 0.8.27 and 0.8.28. The GitHub security advisory GHSA-jw3m-8q7m-f35r, published July 8, 2026, included proof-of-concept steps demonstrating both the hostPath sink (reading /etc/passwd from the node) and the controller sink (reading the service-account token).
What you should do
Upgrade to LLMKube version 0.9.0 or later. The fix, released July 5, 2026, gates local and hostPath model sources and blocks controller server-side request forgery.
What it means
This vulnerability allowed privilege escalation within a Kubernetes cluster. A low-privilege namespace user could extract credentials that grant broader cluster access, potentially compromising workloads in other namespaces or the control plane itself. The CVSS v3.1 base score of 7.7 reflects the high impact on confidentiality and the low attack complexity. Organizations running multi-tenant Kubernetes clusters with LLMKube should audit their Model resources for suspicious absolute paths and review access logs for the model-cache PVC.
Kubernetes tooling around AI workloads has been a recurring weak spot: see our report on the k8s-mcp-server authentication bypass, plus more coverage in our Infrastructure section and the AI Security hub.
Sources
- NVD CVE-2026-68903
- GitHub Advisory GHSA-jw3m-8q7m-f35r
- LLMKube v0.9.0 release notes



