CVE-2026-105753

[CVE catalog]

Severity

Medium · CVSS 6.5CVSS:3.1/AV:N/AC:L/PR:L/UI:N/S:U/C:N/I:N/A:H

Published

Oct 5, 2026, 10:00:00 PM UTC

Updated

Oct 5, 2026, 11:17:01 PM UTC

Source

[NVD] · [CVE record]

EPSS

0.43% · Oct 25, 2026

A flaw was found in vLLM. This vulnerability allows a remote user to cause a Denial of Service (DoS) on the model serving engine. The issue occurs when a multimodal request is rejected after media processing, causing the frontend and backend caches to become desynchronized. A subsequent request reusing the same media identifier triggers an internal assertion failure due to missing cached items, crashing the service. vllm: vllm: Denial of Service via multimodal cache desynchronization A MODERATE severity denial of service vulnerability was found in vLLM as shipped in Red Hat AI Inference Server 3, Red Hat OpenShift AI, and Red Hat Enterprise Linux AI. When a multimodal request is rejected after its media has been rendered and hashed (for example, a chat prompt that exceeds the model's maximum context length), the default mirrored multimodal LRU processor cache records the media hash in the frontend sender cache, but the engine core's receiver cache never receives the payload. A remote client with ordinary access to the inference API can exploit this cache desynchronization by resubmitting the same media item in a later request: the sender cache reports a hit and sends no payload, and the engine core fails an internal assertion ("Expected a cached item"), causing a denial of service on the shared serving instance. Only the availability of the inference service is impacted; there is no impact on confidentiality or integrity, and no code execution or memory corruption is involved. Only the default mirrored LRU multimodal cache configuration is affected. Because this flaw is rated MODERATE with a CVSS 3.1 base score of 6.5, below the threshold at which Red Hat files trackers for moderate-severity flaws in these products, the fix is currently deferred and no updated Red Hat packages are available at this time. Until a fix is delivered, deployments that expose the inference API to untrusted or multi-tenant clients remain at risk of a remote client crashing the service. The CVSS score(s) listed for this vulnerability do not reflect the associated product's status, and are included for informational purposes to better understand the severity of this vulnerability.