GPU VulnDB

Database/AI/ML frameworks & serving

Kubeflow Pipelines: frontend artifact routes skip namespace authorization, exposing other tenants' artifacts

CVSS 6.5CVE-2026-118760AI/ML frameworks & servingcurated

Impact

The Kubeflow Pipelines frontend authenticates the caller at ingress but never checks that they are authorized for the namespace named in an artifact request. Any authenticated user who knows or guesses an artifact location can select another tenant's namespace - or omit the namespace entirely and hit an unguarded direct handler - and read datasets, trained models and pipeline outputs belonging to other teams. On a shared GPU cluster where Kubeflow namespaces are the tenancy boundary, this is a straight cross-tenant read of training data and model weights. The advisory is explicit that it does not give arbitrary Secret disclosure, storage writes, or code execution. Versions 0.5.1 up to 2.17.0.

Who can reach it

Any user authenticated to the Kubeflow Pipelines frontend in a multi-user deployment. No membership in the target namespace is needed.

What to do

Upgrade Kubeflow Pipelines to 2.17.0 and roll the frontend deployment - a standard pod rollout, no node disruption. Where the upgrade has to wait, restrict frontend access and do not rely on KFP namespace isolation for artifact confidentiality.

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.