# CRI-O v1.35.12

2026-11-16 · [Original release notes](https://github.com/cri-o/cri-o/releases/tag/v1.35.12) · [Full summary](https://ratatosk.io/en/releases/cri-o/v1.35.12)

Security 1 (high 1) · Breaking changes 2 · Important fixes 2 · Other changes 3

## Security (1)

### [high] CRI-O linked log volume write vulnerability fixed

- Affects: Not stated in the release notes
- [CVE-2026-106096](https://www.cve.org/CVERecord?id=CVE-2026-106096), [GHSA-pw8r-75cr-c993](https://github.com/cri-o/cri-o/security/advisories/GHSA-pw8r-75cr-c993)

A container could influence CRI-O or conmon writes to host files outside its log directory, compromising host-side integrity and potentially enabling root command execution if a targeted file was later executed. After the fix, containers see linked log volumes as read-only, including through kubelet `subPath` bind aliases, while CRI-O can still write from its host-side view; log collectors must keep offsets and buffers elsewhere.

Security advisory:

> When a container uses the opt-in `link-logs.crio.io` annotation, CRI-O exposes the pod's log directory through a pod-owned emptyDir and intends that view to be read-only. That isolation was not effective, so a container could alter host-side log state that CRI-O and conmon depend on. This gave the container a way to influence writes made by a privileged component against host files outside its own log directory, crossing the container-to-host integrity boundary; where such a file is later executed by the host, it can amount to command execution as root.
> The flaw exists since the feature's introduction in 1.28.0 and is reachable only when the annotation is enabled for the runtime handler; deployments that do not enable it are unaffected. Exploitation depends on pre-existing host state and remains subject to mandatory access control (MAC) policy, which places the attack complexity at high.
> The linked log volume is now mounted read-only in containers even when the container requests write access, including when the volume is reached through kubelet subPath bind aliases. The host-side view stays writable to CRI-O itself so that log linking and SELinux labeling keep working. Users should upgrade to one of:
> This is a behaviour change: log collectors must keep writable state such as offsets and buffers in a separate volume.
> - Do not enable `link-logs.crio.io` for any runtime handler. Deployments that leave it out of `allowed_annotations` are not reachable through a pod.
> - Run the containers of affected pods as an unprivileged UID that cannot become uid 0 (`runAsNonRoot`, no `setuid` escalation, no user-namespace mapping to host root). Kubelet creates the pod and container log directories under `/var/log/pods` as root-owned and not writable by group or other (0755), so such a container cannot create entries in the linked tree. This is conditional hardening: pods whose containers run as root, whether requested or inherited from the image, remain exposed.

https://github.com/cri-o/cri-o/security/advisories/GHSA-pw8r-75cr-c993

## Breaking changes (2)

### Go 1.26.10 minimum for CRI-O builds

- Affects: setups building CRI-O with Go or its static-build compiler (Inferred)

The Go module previously declared Go 1.25.0, and the Nix derivation used Go 1.25; both now use Go 1.26.10.

Evidence: “Update Go to 1.26.10 in `dependencies.yaml` and its existing references in `go.mod` and `nix/derivation.nix` so these builds use the patched toolchain.” https://github.com/cri-o/cri-o/pull/10493

>  - CRI-O's minimum Go version and static-build compiler are now Go 1.26.10. (#10493, @bitoku)

### CRI-O shutdown deadlock fixed

- Affects: CRI-O setups with CNI plugin readiness checks (Inferred)

During node shutdown, in-flight RPCs waiting for CNI readiness could deadlock CRI-O and make it ignore SIGTERM. Shutdown now stops the CNI manager before gracefully stopping gRPC, forces gRPC to stop if graceful shutdown has not completed after 30 seconds, and returns an error when a CNI-readiness wait is canceled by its context.

>  - Fixed a deadlock where CRI-O ignored SIGTERM when in-flight RPCs were blocked waiting for CNI readiness during node shutdown. The CNI manager is now shut down before the gRPC server stops, and a forced stop is scheduled as a fallback after 30 seconds. (#10545, @nispriha)

## Important fixes (2)

### Crash while stopping containers with active exec processes fixed

- Affects: containers with active exec processes

Before using the exec-process PID map while stopping a container, CRI-O now clones it.

>  - Fixed a possible CRI-O crash when stopping containers with active exec processes. (#10529, @openshift-cherrypick-robot)

### CRI-O restores default SELinux labeling for systemd/init containers

- Affects: SELinux-enabled setups running systemd/init containers

With the storage-default `container_t` label, systemd pods on SELinux-enabled nodes could fail to create cgroups with permission denied. An explicitly configured `seLinuxOptions.type` on the container or sandbox still takes precedence.

>  - Fix SELinux labeling so systemd/init containers receive container_init_t again instead of remaining container_t and failing to create cgroups. (#10509, @aksjadha)

## Other changes (3)

- Tracked `golangci-lint` version updated to v2.10.1
- s390x Nix builds disable `zlib` vectorized CRC32
- `ContainerMinMemory` fallback error uses `%d`

---

Summaries are produced by AI, so they can be wrong, and they are provided without warranty.
