---
id: "CVE-2026-126397"
page: "https://infrastack.ai/live/CVE-2026-126397"
kind: "vulnerability"
aliases: ["CVE-2026-126397"]
title: "vLLM: unbounded image patch setting lets remote clients exhaust CPU"
severity: "high"
cvss: 8.7
epss: null
exploited: false
status: "open"
category: "ai-tooling"
products: [{"name": "vllm", "fixed": [], "affected": "", "ecosystem": "pypi"}]
ecosystems: ["pypi"]
tags: ["ai-tooling", "llm-serving", "multimodal", "self-hosted", "vllm"]
not_affected_if: ["managed-hosting", "no-untrusted-clients", "text-only-models"]
published_at: "2026-11-13T10:17:43+00:00"
first_seen_at: "2026-11-13T10:47:19+00:00"
updated_at: "2026-11-13T10:50:22+00:00"
sources: [{"source": "nvd", "url": "https://nvd.nist.gov/vuln/detail/CVE-2026-126397", "first_seen_at": "2026-11-13T10:47:19+00:00"}]
exploit_sources: []
---

# vLLM: unbounded image patch setting lets remote clients exhaust CPU

Link for people: https://infrastack.ai/live/CVE-2026-126397 (this .md is for agents)

vLLM forwards a request-supplied max_patches value to the Hugging Face processor for Cohere2 Vision models with no server-side upper bound. A remote client can send small image requests with large varying values to force quadratic CPU work and deny service to other users.

_Summary, check and fix written by a model from the sources; labelled as such._

## Affected

- pypi:vllm affected range not listed fixed no fixed version listed

## How to check

- Check the installed vLLM version with `pip show vllm` and compare it to the range below.
- Check whether the server serves Cohere2 Vision models that accept image input.

## Fix

- `vllm` `0.31.0` or later (fixed version per the NVD range: before 0.31.0 is affected).

## Timeline

- 2026-11-13T10:47:19+00:00 nvd https://nvd.nist.gov/vuln/detail/CVE-2026-126397

## Documents

- [advisory] Maintainer's advisory: github.com https://github.com/vllm-project/vllm/security/advisories/GHSA-9g3c-x747-c6wr
- [advisory] Third-party advisory: www.vulncheck.com https://www.vulncheck.com/advisories/vllm-0.10.1-before-0.31.0-cohere2-vision-dos-via-max-patches
- [fix] Patch: github.com https://github.com/vllm-project/vllm/commit/f6326f53bda46898a331c2d24500332c285d9a2b

## From the sources

vLLM from 0.10.1 before 0.31.0 forwards a request-controlled mm_processor_kwargs.max_patches value to the Hugging Face processor for Cohere2 Vision models without a server-side upper bound, causing quadratic CPU work during multimodal preprocessing. Remote clients can send small image chat-completion requests with large, varying max_patches values to bypass the LRU cache, exhaust preprocessing workers, and deny service to other clients.

---

_Free to read, with fair use of about 60 pages a minute. For higher limits or API access: sales@infrastack.ai_