vLLM: vllm p2p never connected zmq_exhaustion
vLLM
More for vLLM RSS feed for vLLM
- Severity
- High, CVSS 7.5
- CVSS vector
- CVSS:3.1/AV:N/AC:L/PR:N/UI:N/S:U/C:N/I:N/A:H
- Record
- CVE-2026-94624
- Affects
- vLLM
- Known exploited
- Not listed in the CISA Known Exploited Vulnerabilities catalog
- Published
- Record first published by the source
- September 21, 2026
- Source last modified
- October 27, 2026
Description
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 kv_transfer_params 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.
Description quoted from the NVD record.
Who is affected
vLLM is software that runs AI language models on a server so other programs can send it questions and get answers. This flaw affects vLLM versions through 0.29.0, but only when the feature that moves the model's short-term working data to other machines is turned on and set up with a tiered layout whose second tier connects directly to peer machines. An attacker can send requests pointing to made-up network addresses, which leaves connections open until the server runs out of them. The core of vLLM then crashes and stops answering all requests. This flaw is rated High.
What to do
- Find out who runs vLLM for your organization: the person who installed it, or the IT provider or vendor who manages it.
- No fixed version is named for this flaw yet, so that person should check the vendor advisory linked on this page for a fixed release and install it once it is available.
- That person should also check whether your vLLM setup uses the feature that moves working data to peer machines, since only setups using it are affected.
- If an IT provider or vendor runs vLLM for you, send them this advisory and ask when the update will be done.
Analysis written from the source record as of October 27, 2026 and reviewed before publishing.