Zero Radar
← feed
HighVulnerabilityCVE-2026-114050updated 25 Oct 2026, 13:32 UTC

vLLM engine can be crashed via crafted image/video cache IDs

Anyone who can send requests to a vLLM server can crash its model engine with a few specially crafted image requests. The server then stays down until someone restarts it.

Fixed in vLLM 0.31.0Official, 2 sources
SeverityHighCVSS 7.1 of 10
Being exploited?UnclearNo reports either way yet
Fix available?YesvLLM 0.31.0 · since 25 Oct 2026
How sure is this?Official, 2 sourcesBacked by an official source

How the attack works simplified

  1. 1Client sends imageSets a chosen uuid on the image
  2. 2Frontend cache driftsFiller requests evict the original entry
  3. 3Stale data reusedEngine uses old cached tensor for new image
  4. 4Token mismatch errorEngine process hits an uncaught exception
  5. ResultEngine dies, server needs restart

What happened

Multimodal content parts accept a client-supplied uuid that vLLM uses verbatim as the cache key; the frontend (P0) and engine (P1) caches evict independently and the engine does not validate cache hits, so stale embeddings are substituted, causing a token-placeholder mismatch that kills the EngineCore process.

Why it is dangerous: A single unauthenticated client can take down the whole serving instance with default settings, causing outages for every user of that deployment until it is restarted.

Are you affected?

You are affected if you run vLLM before 0.31.0 with multimodal (image, video or audio) models and expose the OpenAI-compatible API to untrusted clients.

Package / productAffected versionsFixed in
vllm · pip< 0.31.00.31.0

What to do

  1. Upgrade vLLM to 0.31.0 or later
  2. Until upgraded, restrict who can reach the API endpoint and add authentication in front of it
  3. Monitor /health and automatically restart the server if it stops responding

Timeline

  1. 21 Oct 2026, 01:17 UTC
    Publicly disclosed
  2. 21 Oct 2026, 01:17 UTC
    Reported by NVD · NVD
  3. 23 Oct 2026, 20:43 UTC
    Reported by GitHub Advisory · GitHub Advisory

Sources 2 reports, oldest first

Official vendor, government CERT or advisory database

Technical details

ID
CVE-2026-114050 · GHSA-p92p-rxj5-7p2x
CVSS
7.1 / 10
EPSS
0.2% chance of exploitation within 30 days
Disclosed
21 Oct 2026, 01:17 UTC
On Zero Radar
before the radar started
Last change
25 Oct 2026, 13:32 UTC

The plain-language parts and the attack diagram were written by an AI model (anthropic/claude-haiku-5.5) from the sources above. Facts like versions, exploitation and dates come from the sources themselves. How Zero Radar works

vLLM engine can be crashed via crafted image/video cache IDs · Zero Radar