GPU VulnDB

Database/AI/ML frameworks & serving

Jupyter Enterprise Gateway: Jinja2 template injection in KERNEL_* env vars gives RCE in the gateway

CVSS 10.0CVE-2026-44181AI/ML frameworks & servingcurated

Impact

The KERNEL_XXX environment variables a client supplies when requesting a kernel are rendered through Jinja2 into the Kubernetes pod manifest without sanitization, so template expressions execute Python and shell commands inside the Enterprise Gateway service itself. The gateway holds a Kubernetes service account token with the rights to create kernel pods, so the attacker can read cluster secrets and schedule a privileged pod or a pod with a hostPath mount - that is full control of the cluster and of every GPU node in it. On a shared GPU cluster where Enterprise Gateway is the front door for notebook users, this turns a notebook request into cluster-admin. The advisory describes the request path as unauthenticated.

Who can reach it

Anyone who can reach the Enterprise Gateway HTTP endpoint and submit a kernel start request; the CVSS vector and the advisory indicate no authentication is required. If the gateway is only reachable from a tenant network, any tenant on that network qualifies.

What to do

Upgrade Enterprise Gateway to 3.3.0 and restart the gateway service; running kernels are unaffected but in-flight sessions drop. No node reboot or GPU downtime is needed. Until the upgrade, restrict network access to the gateway and treat the gateway service account token as potentially exposed - rotate it and audit for pods created outside the normal kernel pattern.

References

Related entries

All AI/ML frameworks & serving entries

This entry is curated: imported from vendor advisories with machine assistance, not yet individually verified. Confirm against your vendor's advisory before acting, and report anything wrong.