---
id: CVE-2026-34755
title: vLLM is an inference and serving engine for large language models (LLMs)
summary: >-
  vLLM is an inference and serving engine for large language models (LLMs). From
  0.7.0 to before 0.19.0, the VideoMediaIO.load_base64() method at
  vllm/multimodal/media/video.py splits video/jpeg data URLs by comma to extract
  individual JPE…
severity: medium
cvss: 6.5
cvssVector: 'CVSS:3.1/AV:N/AC:L/PR:L/UI:N/S:U/C:N/I:N/A:H'
cwe:
  - CWE-770
vendor: vllm
product: vllm
affected:
  - 'vllm >= 0.7.0, < 0.19.0'
patched:
  - vllm 0.19.0
published: '2026-04-06'
updated: '2026-07-07'
source: NVD
sourceUrl: 'https://nvd.nist.gov/vuln/detail/CVE-2026-34755'
references:
  - url: >-
      https://github.com/vllm-project/vllm/security/advisories/GHSA-pq5c-rjhq-qp7p
    label: security-advisories@github.com
  - url: 'https://access.redhat.com/errata/RHSA-2026:36005'
    label: 0b0ca135-0b70-47e7-9f44-1890c2a1c46c
  - url: 'https://access.redhat.com/errata/RHSA-2026:36006'
    label: 0b0ca135-0b70-47e7-9f44-1890c2a1c46c
  - url: 'https://access.redhat.com/security/cve/CVE-2026-34755'
    label: 0b0ca135-0b70-47e7-9f44-1890c2a1c46c
  - url: 'https://bugzilla.redhat.com/show_bug.cgi?id=2455403'
    label: 0b0ca135-0b70-47e7-9f44-1890c2a1c46c
  - url: >-
      https://security.access.redhat.com/data/csaf/v2/vex/2026/cve-2026-34755.json
    label: 0b0ca135-0b70-47e7-9f44-1890c2a1c46c
  - url: 'https://nvd.nist.gov/vuln/detail/CVE-2026-34755'
  - url: 'https://github.com/vllm-project/vllm/pull/38636'
  - url: >-
      https://github.com/vllm-project/vllm/commit/58ee61422169ce17e08248f8efa1e9df434fe395
  - url: 'https://github.com/vllm-project/vllm'
  - url: >-
      https://github.com/pypa/advisory-database/tree/main/vulns/vllm/PYSEC-2026-144.yaml
  - url: 'https://access.redhat.com/errata/RHSA-2026:59151'
  - url: 'https://access.redhat.com/errata/RHSA-2026:59144'
  - url: 'https://access.redhat.com/errata/RHSA-2026:57390'
  - url: 'https://access.redhat.com/errata/RHSA-2026:57389'
  - url: 'https://access.redhat.com/errata/RHSA-2026:57387'
  - url: 'https://access.redhat.com/errata/RHSA-2026:57380'
tags:
  - nvd
  - osv
  - pip
epss: 0.00843
epssPercentile: 0.5611
ingestedAt: '2026-07-07T12:52:59.574Z'
aliases:
  - GHSA-pq5c-rjhq-qp7p
  - PYSEC-2026-144
ecosystem: pip
---

## Overview

vLLM is an inference and serving engine for large language models (LLMs). From 0.7.0 to before 0.19.0, the VideoMediaIO.load_base64() method at vllm/multimodal/media/video.py splits video/jpeg data URLs by comma to extract individual JPEG frames, but does not enforce a frame count limit. The num_frames parameter (default: 32), which is enforced by the load_bytes() code path, is completely bypassed in the video/jpeg base64 path. An attacker can send a single API request containing thousands of comma-separated base64-encoded JPEG frames, causing the server to decode all frames into memory and crash with OOM. This vulnerability is fixed in 0.19.0.

## Affected

- `vllm >= 0.7.0, < 0.19.0`

## Remediation

Upgrade past the affected range:

- `vllm 0.19.0`

## Package advisory (CVE-2026-34755)

Affected packages:

- `vllm >= 0.7.0, < 0.19.0`

Patched in:

- `vllm 0.19.0`

Source: https://osv.dev/vulnerability/GHSA-pq5c-rjhq-qp7p