GPU VulnDB

Database/AI/ML frameworks & serving

Kubeflow Training Operator: PyTorchJob update webhook lets users bypass immutability checks

CVSS 5.4CVE-2026-83494AI/ML frameworks & servingcurated

Impact

An incorrect object cast in the PyTorchJob ValidateUpdate webhook means fields that should be immutable can be changed after a job is created. Per the advisory, a user can detach a queued job from its quota management system or take over controller ownership of a running job. On a shared GPU cluster, that lets one tenant get GPU capacity outside its quota or disrupt another team's training run. The vendor rates the main impact as denial of service. Nothing in the record points to code execution or reading data.

Who can reach it

An authenticated cluster user who is allowed to create or update PyTorchJob objects, over the Kubernetes API. Network-reachable, low privileges required.

What to do

Update the training-operator once Red Hat ships the OpenShift AI fix (track the Red Hat CVE page and bugzilla 2526348). The record names no fixed version. Patching the operator means rolling its controller and webhook deployment, and running training jobs do not need a drain. Until then, limit who can update PyTorchJob objects with RBAC, and audit quota and ownership fields on queued and running jobs.

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.