Cloudanix VulnDB

CVE-2026-94624

High · 7.5

CVSS

7.5

EPSS

0.4%

Published

Sep 21, 2026

CISA due date

—

Cloudanix analysis

Next patch cycle SSVC: Track

A configuration issue in certain AI software lets attackers feed bad network addresses that eventually exhaust resources and shut down the entire inference system.

Exploitation
Remote Unauthenticated
Attacker needs
None
Complexity
Low
Impact
Denial Of Service
Component
Application
Class
Denial of service
Remediation
Upgrade
Effort
Medium
Patch risk
Low

What it actually does

The vulnerability allows an attacker to supply arbitrary remote host and port values in kv_transfer_params when OffloadingConnector is configured with TieringOffloadingSpec and a peer-to-peer secondary tier, creating unreachable peer sessions that retain ZeroMQ sockets until the context quota is exhausted and triggering an uncaught ZMQError that crashes EngineCore.

Why it matters to the business

Successful exploitation causes complete interruption of inference services in affected Red Hat AI products, resulting in operational downtime for machine learning workloads with low urgency given the absence from KEV and stable low EPSS.

How an attack unfolds

An attacker sends crafted kv_transfer_params containing unreachable remote endpoints to a vLLM instance with P2P KV offloading enabled, forcing repeated creation of stalled ZeroMQ sessions that exhaust the socket quota and crash the inference engine.

How to fix it

Upgrade rhoai/odh-kserve to agent-rhel9:1791468381 and rhoai/odh-kserve-storage to initializer-rhel9:1790958549 as provided in RHSA-2026:84734; consider disabling P2P KV offloading when not required.

How to confirm the fix worked

Confirm package versions match the fixed releases listed in RHSA-2026:84734 and verify that EngineCore remains stable when TieringOffloadingSpec is exercised with invalid peer parameters.

Typical attack path

  1. Locate a vLLM deployment with OffloadingConnector and TieringOffloadingSpec enabled
  2. Submit inference requests containing malicious kv_transfer_params with arbitrary unreachable host and port values
  3. Trigger repeated allocation of unreachable peer sessions that retain ZeroMQ sockets
  4. Exhaust the ZeroMQ context quota
  5. Induce uncaught ZMQError to crash EngineCore and halt all inference

How to detect it

Inspect vLLM and EngineCore logs for repeated ZMQ socket allocation failures or uncaught ZMQError exceptions that coincide with inference requests containing unexpected remote host or port values.

What your security team should do

  • Apply the Red Hat security updates from RHSA-2026:84734 without delay
  • Disable P2P KV offloading in configurations where it is not essential
  • Review inference request logs for anomalous kv_transfer_params values

The record

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.

Affected packages

Ecosystem Package Introduced Fixed

Related vulnerabilities