---
id: CVE-2026-90554
title: >-
  vLLM versions >=0.10.2 and <0.28.0 do not apply any audio decode-size or
  duration limit when extracting audio from video input for NanoNemotronVL
  models
summary: >-
  vLLM versions >=0.10.2 and <0.28.0 do not apply any audio decode-size or
  duration limit when extracting audio from video input for NanoNemotronVL
  models. In nano_nemotron_vl.py, _extract_audio_from_videos calls
  load_audio_pyav(BytesIO(vi…
severity: medium
cvss: 6.2
cvssVector: 'CVSS:3.1/AV:L/AC:L/PR:N/UI:N/S:U/C:N/I:N/A:H'
cwe:
  - CWE-400
  - CWE-770
vendor: vllm
product: vllm
affected:
  - 'vllm >= 0.10.2, < 0.28.0'
patched:
  - vllm 0.28.0
published: '2026-09-12'
updated: '2026-09-24'
sourceUpdated: '2026-09-24T20:28:13.590'
source: NVD
sourceUrl: 'https://nvd.nist.gov/vuln/detail/CVE-2026-90554'
references:
  - url: >-
      https://github.com/vllm-project/vllm/security/advisories/GHSA-936p-m5pv-vvjf
    label: disclosure@vulncheck.com
  - url: >-
      https://www.vulncheck.com/advisories/vllm-before-0.28.0-denial-of-service-via-audio-extraction
    label: disclosure@vulncheck.com
  - url: >-
      https://security.access.redhat.com/data/csaf/v2/vex/2026/cve-2026-90554.json
  - url: 'https://access.redhat.com/security/cve/CVE-2026-90554'
  - url: 'https://bugzilla.redhat.com/show_bug.cgi?id=2532572'
  - url: 'https://www.cve.org/CVERecord?id=CVE-2026-90554'
  - url: 'https://nvd.nist.gov/vuln/detail/CVE-2026-90554'
tags:
  - nvd
  - cve.org
  - csaf
  - vex
  - red-hat
epss: 0.0016
epssPercentile: 0.04464
ssvc:
  exploitation: none
  automatable: 'no'
  technicalImpact: partial
  timestamp: '2026-09-14T16:14:37.760892Z'
ingestedAt: '2026-09-14T15:23:07.479Z'
---

## Overview

vLLM versions >=0.10.2 and <0.28.0 do not apply any audio decode-size or duration limit when extracting audio from video input for NanoNemotronVL models. In nano_nemotron_vl.py, _extract_audio_from_videos calls load_audio_pyav(BytesIO(video_bytes)) without the max_duration_s or max_decode_bytes parameters, so neither VLLM_MAX_AUDIO_DECODE_DURATION_S nor VLLM_MAX_AUDIO_DECODE_BYTES is enforced (unlike the direct audio upload path in AudioMediaIO). When a NanoNemotronVL model is served with use_audio_in_video=True, an attacker who supplies a small, highly compressed video as multimodal input can force the server to allocate gigabytes of memory during audio decoding, resulting in a denial of service. Fixed in vLLM 0.28.0.

## Affected

- `vllm >= 0.10.2, < 0.28.0`

## Remediation

Upgrade past the affected range:

- `vllm 0.28.0`

## Vendor advisories

- **Red Hat VEX** · Moderate · affected: Red Hat AI Inference Server, Red Hat Enterprise Linux AI (RHEL AI) 3, Red Hat OpenShift AI (RHOAI) · no fix planned: Red Hat AI Inference Server, Red Hat Enterprise Linux AI (RHEL AI) 3, Red Hat OpenShift AI (RHOAI) · updated 2026-09-24 · [vex](https://security.access.redhat.com/data/csaf/v2/vex/2026/cve-2026-90554.json)