GPU VulnDB

Database/AI/ML frameworks & serving

ktransformers: unauthenticated pickle deserialization on the scheduler ZMQ socket gives RCE as the server process

CVSS 9.3CVE-2026-63767AI/ML frameworks & servingcurated

Impact

ktransformers is an LLM inference framework run on GPU nodes for local/offloaded serving. Its SchedulerServer binds a ZMQ ROUTER socket to all interfaces and unpickles whatever arrives, so anyone who can reach that port runs arbitrary commands as the serving process - no credentials needed. On a GPU node that process typically holds the model weights, the KV cache and a GPU context, so the attacker inherits read access to loaded models and tenant prompts and a foothold on a node that is expensive to drain. If the inference user can reach the container runtime socket or shared cluster storage, this is a pivot off the node rather than a single-process compromise.

Who can reach it

Anyone with network reach to the SchedulerServer ZMQ port, which is bound to all interfaces by default. No authentication is required. In practice this is any tenant pod on the same network or anyone on a flat cluster/management network; it is internet-reachable if the node is exposed.

What to do

Update past 0.6.3 to a build containing commit def0f93 (fixed via PR 2091) and restart the scheduler process - the serving daemon must be bounced, which drains in-flight requests. Until then, bind the scheduler socket to localhost or restrict the port with host/network policy so only the intended inference workers can reach it. The record names a fix commit, not a tagged release.

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.