CVE-2026-90554

What it is

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.

At a glance

Exploitability

Can be exploited locally on the machine without logging in. High risk of service disruption.

Exploit odds

0.2% chance of exploitation in the next 30 days (EPSS).

Fix

A patch, advisory or fixed version is referenced. Upgrade or apply it.

Timeline

1. 2026-09-12 13:16 UTC PublishedDisclosed publicly on CVE.org and NVD
2. 2026-09-24 20:28 UTC Last updatedMost recent change to the record

Details

Published

2026-09-12 13:16 UTC

Last modified

2026-09-24 20:28 UTC

Exploit probability

0.2% · EPSS, next 30 days

Weakness

CWE-400 · Uncontrolled resource consumption

Record status

Analyzed

Affected

• Vllm < 0.28.0

Attack profile CVSS 4.0 · 6.9

Attack vector

Local

Complexity

Low

Attack requirements

None

Privileges needed

None

User interaction

None

Confidentiality

None

Integrity

None

Availability

High