GHSA-m38g-vww2-mvgx

Summary

A vulnerability in the Linux kernel's algifaead subsystem (CVE-2026-31431, "copy.fail") allows an unprivileged container workload to corrupt arbitrary file page-cache pages via the AFALG crypto interface and splice(). On Talos Linux, this vulnerability can be chained into a complete node compromise: an attacker who can schedule a pod on a worker node can, without any elevated Kubernetes permissions, achieve arbitrary code execution as root on the host (by poisoning a binary inside a privileged pod, or poisoning a binary which runs with elevated privileges like a CNI binary), access host filesystem, including node secrets.

The exploit does not require kernel debugging, race conditions, or any prior privileges beyond the ability to create a pod.

Impact

An attacker with the ability to deploy a Kubernetes pod on an affected node can:

1. Corrupt the page-cache of /usr/sbin/nft in the containerd snapshot layer shared between the attacker's pod and the kube-proxy DaemonSet. Because containerd reuses XFS page-cache pages across overlayfs mounts sharing the same lower layer, the corruption is immediately visible to all containers using that image layer — including privileged system DaemonSets.
2. Execute arbitrary code inside kube-proxy — a privileged DaemonSet running on every node with all Linux capabilities (privileged: true) and host network access — the next time kube-proxy invokes nft as part of its nftables reconciliation loop (typically within seconds).
3. At this point, an attacker achieved code execution inside a privileged pod, which allows to escape to the host.
4. Same attack can be planted by infiltrating other binaries running as privileged, for example a CNI plugin.

Patches

Upgrade to Talos v1.13.0 or Talos v1.12.7 which ships Linux kernel 6.18.25. The kernel fix for CVE-2026-31431 (algif_aead in-place optimization revert) was committed upstream in Linux 6.18.22 and is included in all Talos releases from v1.13.0 and Talos 1.12.7 onwards.

Workarounds

There are multiple workarounds available based on the situation, but upgrading is strongly recommended.

Option 1 - Change kernel arguments

Add a kernel argument with initcall_blacklist=algif_aead_init by upgrading Talos to the same version.

Note: this either requires setting machine.kernel.extraKernelArgs if using BIOS based boot or upgrading with a new image from factory/imager generated image by setting the extra kernel args. See Boot Assets

Option 2 - Deploy all workload pods with a seccomp profile denying creating AF_ALG socket creation

patch.yaml

machine:
  seccompProfiles:
    - name: copy-fail-block.json
      value:
        defaultAction: SCMP_ACT_ALLOW
        syscalls:
          - names:
              - socket
            action: SCMP_ACT_ERRNO
            args:
              - index: 0
                value: 38
                op: SCMP_CMP_EQ

Apply this patch to all machines in the cluster and set this for all the pod spec:

...
spec:
      securityContext:
        seccompProfile:
          type: Localhost
          localhostProfile: profiles/copy-fail-block.json

Option 3 - Block the syscall in runtime with a eBPF program

See copy-fail-blocker, this can be applied to a running system without a reboot, but it has to run before any other workloads are scheduled after a reboot.

Resources

• https://copy.fail/
• https://xint.io/blog/copy-fail-linux-distributions
• https://github.com/theori-io/copy-fail-CVE-2026-31431
• https://github.com/Percivalll/Copy-Fail-CVE-2026-31431-Kubernetes-PoC