CVE-2026-94624: vLLM through 0.29.0 Denial of Service via Unbounded P2P KV Offloading Sessions
vLLM through 0.29.0 contains a denial of service vulnerability in P2P KV offloading when OffloadingConnector is configured with TieringOffloadingSpec and a peer-to-peer secondary tier. Attackers can supply arbitrary remote host and port values in kvtransferparams to create unreachable peer sessions that retain ZeroMQ sockets until the context quota is exhausted, causing an uncaught ZMQError that crashes EngineCore and stops all inference.
Affected Software
Event History
Frequently Asked Questions
Which deployments are exposed to this issue?
Deployments are exposed when vLLM uses P2P KV offloading through an OffloadingConnector configured with a TieringOffloadingSpec and a peer-to-peer secondary tier. The affected versions are vLLM through 0.29.0.
What does an attacker need to do to trigger the denial of service?
An attacker must be able to supply kv_transfer_params containing arbitrary remote host and port values. Unreachable peers create sessions that retain ZeroMQ sockets until the context quota is exhausted.
What is the operational impact after exploitation?
When the ZeroMQ context quota is exhausted, an uncaught ZMQError crashes EngineCore. This stops all inference handled by that EngineCore instance.
What configuration change can reduce exposure if patching is not immediately possible?
Avoid configuring P2P KV offloading with a TieringOffloadingSpec and a peer-to-peer secondary tier. Also prevent untrusted parties from supplying arbitrary remote host and port values through kv_transfer_params.