What containerd/containerd shipped

This week: ignore_descriptor_urls option for pulls, a sandboxed runc shim with no pause container, console kept open when stdin closes, Windows bundle and sparse file fixes, and request body cleanup on auth failures backported across releases.

This week in containerd: pulls gain an ignore_descriptor_urls option with a descriptor URL request counter and sanitized logging, and a sandboxed runc shim now implements the Sandbox API so a pod needs no pause container or pause image. Other changes keep the console open when stdin closes on a terminal process, avoid symlinks for the bundle work directory on Windows so task creation no longer needs administrator rights, and set the sparse file attribute before mkfs. The docker remote now closes request bodies on authorization failures, backported to release branches.

Weekly digest: runc shim no longer blocks State on in-flight Delete, gRPC updated to v1.85.0 to fix a transport deadlock, SELinux pod labels allocated in the CRI layer, AF_UNIX socket support on Windows, and overlay no longer overrides a configured index mount option.

This week in containerd: the runc shim no longer holds its lock across full Delete cleanup, so State calls are not blocked behind slow deletes. The vendored gRPC is updated to v1.85.0 to fix a transport deadlock under concurrent RPCs. SELinux pod labels are now allocated and tracked in the CRI layer, and runtime handlers for image pulls resolve to their configured platform and snapshotter. The overlay snapshotter no longer overrides a user-configured index mount option, backported across release branches. The dialer now accepts AF_UNIX sockets on Windows next to named pipes, the Rust runc shim runs in CI, NRI is updated to 0.12.5, and legacy gRPC tracing is disabled in the Makefile.

This week: Go 1.27.3 with a crypto/x509 security fix on main and release/2.4, runhcs v0.15.0, a Docker pusher body leak fix, Copy accepting -1 as unknown size, and CRI pull progress on request events.

This week in Containerd: Go was updated to 1.27.3, which carries a crypto/x509 security fix, and the update reached release/2.4. runhcs moved to v0.15.0 on main and was backported to the 2.3 and 2.4 branches. The Docker remote pusher now closes the upload request body when authorization fails, fixing a leak when authorization does not succeed. core/content Copy again supports -1 as an unknown size, backported to 2.3 and 2.4. CRI records pull progress on request events so empty-body responses and brief active-request changes are observed. Upgrade tests were adjusted on 2.3 and 2.4, and flaky CI and documentation anchors were fixed.

Three patch releases shipped this week: 2.2.10, 2.3.7 and 2.4.2, all with security patches. Also in: an opt-in erofs fsync_on_commit option, terminal I/O released on delete for runc-v2, a runtime_platforms migration fix in CRI, and gVisor runsc integration buckets.

This week containerd shipped patch releases on three lines: 2.2.10, 2.3.7 and 2.4.2, each containing fixes, updates and security patches including CVE-2026-95839. Work also landed on layer durability for the erofs snapshotter: a new fsync_on_commit option, off by default, fsyncs a layer before its snapshot is committed so a power loss during or shortly after a pull cannot leave corrupt blobs or missing directories. The runc-v2 shim now releases terminal I/O when a terminal process is deleted, the CRI runtime_platforms migration was fixed for release 2.4, and integration buckets for gVisor runsc were added along with fixes for nil spec and runtime option test flakes.

CRI stops following symlinks for container log files. Unknown runtime handlers in PullImage are now rejected. Multipart image fetch gets stability fixes against hangs and corruption. Test fixes backported to 2.0, 2.2 and 2.3.

This week in containerd: container log files are no longer read through symlinks, hardening against link-based log access. CRI PullImage now rejects unknown runtime handlers, matching the configured handlers in the runtime service; an empty handler still behaves as before. Multipart image fetch received stability fixes that avoid hangs and data corruption. Integration test fixes for TestContainerExecLargeOutputWithTTY landed in the 2.0, 2.2 and 2.3 branches, and the critest OOMKilled testcase is skipped for systemd cgroup in 2.0 and 2.2. A planned date was added for release 2.5.

This week: Go 1.27.0 minimum for K8s 1.38 compatibility, reference parser security fix, stricter seccomp socket rules, and proxy snapshotter timeouts across multiple releases.

Containerd this week shipped security and compatibility improvements across releases. Go minimum version bumped to 1.27.0 for Kubernetes 1.38 support. Reference parser now properly rejects path traversal attempts in image references. Seccomp profile switched to explicit socket domain allow-listing. Proxy snapshotter gains configurable timeout support for improved stability.

Containerd patch releases: 1.7.36, 2.0.13, 2.2.9, 2.3.6, 2.4.1. Security fix CVE-2026-53493 and CRI/image storage improvements.

We released patch updates across containerd versions 1.7.36, 2.0.13, 2.2.9, 2.3.6, and 2.4.1 this week. These releases address security concern CVE-2026-53493 alongside important fixes for task leaks during failed container starts, SELinux relabeling failures, image layer fetching, and command-line argument parsing. runc was also updated to v1.5.2 with a kernel bug workaround.

Containerd 2.4.0 and API 1.12.0 are now available. This non-LTS release includes bug fixes for registry uploads, CLI flags, and container isolation. See release notes for full details.

Containerd 2.4.0 and API 1.12.0 releases are now available. As a regular non-LTS release, 2.4.0 introduces new features with a shorter support window for early adopters. Key fixes include content upload retry handling, CLI flag parsing in subcommands, and default masking of thermal interrupt information in containers. The API release adds sandbox controller update propagation and media type tracking in content events.