GPU VulnDB

Database/Kernel, userspace & hypervisor

Linux kernel: out-of-bounds write in watch_queue filters lets a local user gain root

CVE-2022-0995Kernel, userspace & hypervisorKnown exploitedcurated

Impact

An out-of-bounds write in the watch_queue notification subsystem lets a local user overwrite kernel state, giving privileged access or a crash. Unlike most kernel CVEs in obscure drivers, this one lives in a general-purpose subsystem that unprivileged processes can drive on distributions that enable CONFIG_WATCH_QUEUE, and it is reachable from inside an ordinary container unless the syscall path is filtered - so on a multi-tenant GPU node any pod that can open a pipe and set a watch filter has a route to host root, and from there to every other tenant's GPUs, model weights and fabric credentials. Public write-ups exist and CISA lists it as exploited in the wild, so treat it as weaponised rather than theoretical.

Who can reach it

Local unprivileged user on the host, including a process inside a container that has not had the watch_queue path blocked by seccomp. No special capability is required beyond an ordinary shell or pod.

What to do

Take the distribution kernel that contains commit 93ce93587d36 (fixed before 5.17 final; every maintained distro shipped it in 2022). Applying it means rebooting each node, so drain GPU workloads and checkpoint long training runs first - this is a rolling node-reboot campaign, not a service restart. Where a reboot cannot be scheduled immediately, a seccomp profile that blocks the watch_queue ioctl, or building kernels without CONFIG_WATCH_QUEUE, removes the reachable path.

References

Related entries

All Kernel, userspace & hypervisor entries

This entry is curated: imported from vendor advisories with machine assistance, not yet individually verified. Confirm against your vendor's advisory before acting, and report anything wrong.