containerd patches a runAsNonRoot bypass across every supported branch

The containerd maintainers shipped coordinated patches on May 20 — v2.3.1, v2.2.4, v2.0.9, and v1.7.32 — fixing CVE-2026-46680, a moderate-severity advisory that lets a crafted image bypass a pod’s runAsNonRoot restriction.

What the advisory covers

GHSA-fqw6-gf59-qr4w describes numeric User directives in OCI specs that exceed the 32-bit integer range. When the value falls outside that range, containerd parses it as a username instead, then resolves it against the image’s /etc/passwd. An image whose passwd file maps that string to UID 0 lands the container as root, even when the pod spec asked for a non-root numeric user. The advisory lists containerd v1.7.27+ and v2.0.4+ as affected.

Other notable fixes

• AF_ALG is now blocked in the default seccomp socket policy, removing a kernel crypto API path from container processes.
• The overlayfs rebase capability is disabled when running in a user namespace, which had been breaking layer extraction.
• The sandbox task API now handles non-runc runtimes correctly; task fields in Runc options are deprecated.
• Metadata and mount plugin boltdb files now close cleanly on server shutdown.
• The transfer plugin no longer errors out when EROFS differ is configured without mkfs.erofs on the host.

Mitigations short of upgrading

The advisory lists restricting image imports to trusted sources, enforcing a numeric runAsUser in Pod security contexts, or upgrading to Kubernetes 1.34+ — which validates runAsUser independently of the runtime — as workarounds for clusters that cannot patch immediately. The 2.x branches also pick up Go 1.26.3 and bump the containerd API to v1.11.1.

Source: containerd security advisory GHSA-fqw6-gf59-qr4w — May 20, 2026.