
containerd disclosed CVE-2026-53493 last week: a crafted OCI image index can burn unbounded CPU and memory during PullImage. The bug sits in descriptor-graph traversal. Nested or heavily fanned-out indexes had no real depth or breadth limits, and duplicate refs were not deduped well enough. Fixed in 1.7.36, 2.0.13, 2.2.9, 2.3.6, and 2.4.1. The stall happens before the container even starts. If a node can pull from an untrusted registry, or a tenant can set an arbitrary image, this is a node availability problem, not just a bad pod. Patch containerd on every node, restrict pulls to allowlisted registries, and treat long ContainerCreating plus runtime CPU spikes as a pull-path incident until you confirm the version. Image admission and short-lived pull credentials matter here more than anything you do after the container starts.
Post summary
The text discloses CVE-2026-53493 in containerd, a resource exhaustion vulnerability via crafted OCI image index, lists fixed versions, and advises patching nodes and restricting image pulls.

