CVE-2026-90554
Received Received - Intake

Memory Exhaustion in vLLM via Unlimited Audio Decoding

Publication date: 2026-09-12

Last updated on: 2026-09-12

Assigner: VulnCheck

Description
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.

CVSS Scores

EPSS Scores

Probability:
Percentile:

Meta Information

Published
2026-09-12
Last Modified
2026-09-12
Generated
2026-09-12
AI Q&A
2026-09-12
EPSS Evaluated
N/A
NVD
CVE-2026-90554
EUVD
EUVD-2026-76872

Affected Vendors & Products

Showing 1 associated CPE
Vendor Product Version / Range
vllm vllm From 0.10.2 (inc) to 0.28.0 (exc)

Helpful Resources

Exploitability

CWE
CWE Icon
KEV
KEV Icon
CWE ID Description
CWE-400 The product does not properly control the allocation and maintenance of a limited resource.

Attack-Flow Graph

AI Quick Actions

Instant insights powered by AI
Executive Summary

This vulnerability affects vLLM versions between 0.10.2 and 0.28.0. It allows an attacker to upload a small, highly compressed video file that forces the server to allocate excessive memory during audio extraction, potentially causing a denial of service by crashing the server.

Detection Guidance

Check if your vLLM version is between 0.10.2 and 0.27.x by running: vllm --version. If vulnerable, monitor for excessive memory usage during video processing tasks, especially with NanoNemotronVL models. Look for crashes or high memory consumption when handling video inputs.

Impact Analysis

If you use vLLM versions 0.10.2 to 0.27.x with NanoNemotronVL models and enable video audio extraction, an attacker could exploit this to crash your server by consuming gigabytes of memory, leading to service disruption.

Compliance Impact

This vulnerability primarily impacts system availability by enabling denial-of-service attacks through excessive memory consumption during audio extraction from video inputs. For compliance standards like GDPR or HIPAA, which require maintaining system availability and protecting against disruptions, this vulnerability could lead to violations if exploited. Unplanned downtime may result in unauthorized access opportunities or failure to meet service level agreements.

Mitigation Strategies

Upgrade vLLM to version 0.28.0 or later immediately. If upgrading is not possible, disable audio extraction from video inputs for NanoNemotronVL models by setting use_audio_in_video=False. Restrict access to video upload endpoints to trusted users only.

Chat Assistant

Ask questions about this CVE
Hi! I’m here to help you understand CVE-2026-90554. Ask me anything about the vulnerability, its impact, or mitigation strategies.
0/70

EPSS Chart