GPU VulnDB

Database/AI/ML frameworks & serving

Mercury: tokenless deployments let anonymous clients start kernels and execute arbitrary Python

CVSS 9.8CVE-2026-111058AI/ML frameworks & servingcurated

Impact

In default tokenless Mercury deployments before 3.2.6 the endpoint blocklist does not cover /api/kernels, and the websocket wrapper tracked idle activity without authorizing executable messages. An anonymous client can therefore create a Jupyter kernel and send execute_request messages containing arbitrary Python. That is remote code execution as the Mercury service account on whatever node hosts it, and because Mercury is used to publish notebooks to colleagues it is frequently reachable from a wider internal network than the notebooks themselves. On a GPU node the kernel it spawns inherits the pod's device access, so the attacker gets the accelerators, the mounted datasets and any credentials in the environment.

Who can reach it

Any client that can reach the Mercury HTTP/websocket port. No authentication is needed in the default tokenless configuration.

What to do

Upgrade to Mercury 3.2.6 and restart the service; the fix is in commit 73a40ce. Deployments that cannot upgrade immediately should be put behind an authenticating proxy and taken off any broadly reachable interface, and any node that ran an exposed instance should be checked for kernels and processes started by the service account.

References

Related entries

All AI/ML frameworks & serving 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.