GPU VulnDB

Database/Kernel, userspace & hypervisor

Linux kernel net/rds: unprivileged container reads every RDS socket and connection on the host

UnscoredCVE-2026-97476Kernel, userspace & hypervisorcurated

Impact

The RDS_INFO_* getsockopt handlers walked file-scope global lists without filtering by the caller's network namespace, and neither rds_create() nor rds_info_getsockopt() required any capability. A process in a fresh user namespace plus network namespace - which is what a container is - could therefore read the bound address and socket inode of every RDS socket on the host, the peer addresses of incoming messages, and the peer addresses plus TCP and RDS sequence numbers of every RDS connection. On a multi-tenant GPU node that is a namespace-crossing information leak about the host's storage and interconnect traffic; the sequence numbers in particular are material for anyone trying to interfere with an rds-tcp connection. Read-only: no write, no corruption, no code execution. Exposure requires the rds and rds_tcp modules to be loaded, which most GPU nodes do not do unless they use RDS - blacklisting them removes the issue entirely.

Who can reach it

Any local unprivileged user or container process that can create an AF_RDS socket, on a host where the rds module is loaded. No capabilities and no authentication beyond local access are required.

What to do

Take the stable kernel fix that filters each handler by the caller's netns and reboot the node - a drain-and-reboot per node on the usual kernel-update cadence. Where the kernel update cannot be scheduled, confirm whether RDS is actually in use; if it is not, blacklist the rds and rds_tcp modules, which closes the path without a reboot window of its own.

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.