GPU VulnDB

Database/Container, Kubernetes & orchestration

CRI-O: restoring a malicious checkpoint keeps the source security context and bypasses pod policy

CVSS 8.8CVE-2026-92574Container, Kubernetes & orchestrationcurated

Impact

A tenant who can create a pod from a checkpoint image they control gets a process that keeps the credentials, Linux capabilities, no_new_privs flag and seccomp state baked into the checkpoint instead of the ones the destination pod spec asks for. On a shared GPU cluster that is a direct escape from the security context an operator thought it was enforcing: the restored container can run with capabilities and a seccomp profile that admission policy would never have admitted, on a node that also hosts other tenants' GPU workloads. Red Hat scores it 8.8 with confidentiality, integrity and availability all high. CRI-O 1.34 and later are affected upstream; downstream OpenShift from 4.17 onward.

Who can reach it

An authenticated cluster user with permission to create a pod, on a cluster where CRI-O checkpoint/restore is enabled, who can supply a checkpoint image of their choosing. No node access and no host foothold needed.

What to do

No fixed release exists yet - Red Hat states the fixes are merged to supported branches but not shipped. Until a build lands, the practical mitigation is to turn off checkpoint/restore support in CRI-O (it is not enabled by default in every distribution) and to restrict who may create pods from checkpoint images, for example by admission policy on the image source. Both are configuration changes; disabling the feature means editing the CRI-O config and restarting the runtime on each node, which evicts the pods on that node.

References

Related entries

All Container, Kubernetes & orchestration 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.