← All advisories
CVE-2026-126397highDenial of service2026-11-13

CVE-2026-126397: vLLM lets clients pass a huge max_patches to Cohere2 Vision preprocessing and tie up the server, fixed in 0.31.0

vLLM 0.10.1 up to 0.31.0 forwards a request-controlled mm_processor_kwargs.max_patches value to the Hugging Face processor when serving Cohere2 Vision models, with no server-side cap. A small image plus a large max_patches causes quadratic CPU work in preprocessing, and varying the value skips the cache, so any client of the chat-completions endpoint can starve other users. Upgrade to 0.31.0.

01.

What it is

The OpenAI-compatible server accepts mm_processor_kwargs from the request and passes max_patches straight to the Cohere2 Vision processor.

The processor enumerates candidate tile aspect ratios in two nested loops bounded by max_patches, so CPU work grows with the square of the value before normal inference limits apply.

Changing max_patches on each request avoids the helper's small LRU cache, keeping multimodal preprocessing workers busy and raising latency for everyone. No code execution or data exposure was shown.

Only deployments serving Cohere2 Vision models are affected. GitHub advisory GHSA-9g3c-x747-c6wr lists >= 0.10.1, < 0.31.0 as affected and 0.31.0 as patched.

02.

How to check

  1. Run pip show vllm and see whether the version is from 0.10.1 up to (not including) 0.31.0.
  2. Check whether you serve a Cohere2 Vision model and whether untrusted clients can call /v1/chat/completions.
03.

How to fix

  1. Upgrade vLLM to 0.31.0 or later.
  2. Until you can upgrade, put the endpoint behind authentication and per-client rate limits, or strip mm_processor_kwargs from incoming requests at your gateway.
pip show vllm | grep -i ^version
pip install -U 'vllm>=0.31.0'
04.

Sources

More advisories for this tool
The weekly email

One email a week. Only what changed.

Every new security hole in AI coding tools, MCP servers, app builders and the stacks they generate (Next.js, Supabase, Node, PHP), each with how to check your app and the fix. No roundups of things you already read.

Free, one email a week. Read the latest issue first if you like. One-click unsubscribe in every email. We never sell or share addresses.