GPU VulnDB

Database/Container, Kubernetes & orchestration

containerd: malformed subject-annotated manifest digest panics the daemon on image pull

CVSS 7.1CVE-2026-95839Container, Kubernetes & orchestrationcurated

Impact

Pulling or importing an image whose manifest has a child descriptor with the io.containerd.manifest.subject annotation and a malformed digest triggers an unrecovered panic that crashes containerd. A default CRI PullImage request reaches this path. In Kubernetes, kubelet retries keep triggering the crash, so container operations on the worker node are disrupted for as long as the pod that references the image stays scheduled. On a GPU node that means jobs cannot start or be managed, and the node loses work it is paid to run.

Who can reach it

Anyone who can get a node to pull a crafted image: a tenant or user able to create pods that reference an image of their choosing, or whoever controls a registry or image the cluster pulls. Low privileges are needed and no host access.

What to do

Upgrade containerd to 2.2.10, 2.3.7 or 2.4.2 (affected from 2.2.0) and restart the containerd daemon on each node. Running containers normally survive a containerd restart, but schedule the restart per node. Until patched, restricting which registries pods can pull from limits who can trigger it.

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.