GPU VulnDB

Database/Kernel, userspace & hypervisor

OpenSSH ssh-agent: locking bypass lets a forwarded remote session add tokens and use keys

CVE-2026-73281Kernel, userspace & hypervisorcurated

Impact

Operations that ssh-agent was supposed to confine to the local machine - adding tokens and using loaded keys - can be driven from the far end of a forwarded agent socket, because agent locking and the session-bind extension do not interact correctly. On a GPU fleet this matters wherever operators forward their agent into nodes: a hostile or already-compromised node, or another root user on it, can exercise the engineer's keys against the rest of the estate rather than only the box they logged into. Locking the agent, which is the usual habit for exactly this risk, does not stop it. The record scores it 3.5 with scope change and no confidentiality loss; the practical exposure is authentication reach, not key extraction.

Who can reach it

Whoever can reach a forwarded ssh-agent socket on a remote host - a node operator, root on that node, or a co-tenant that obtained the socket. Requires that the user forwarded their agent to that host; no separate credential is needed beyond socket access.

What to do

Upgrade OpenSSH to 10.5 on the machines where ssh-agent runs (operator workstations and bastions, not just the GPU nodes) and restart the agents - a logout/login or fresh agent process picks it up, no node reboot. As an interim measure, stop using ForwardAgent into shared or tenant-adjacent nodes and use ProxyJump instead; that removes the exposure without a package upgrade.

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.