vLLM: VLLM versions >=0.10.2 and <0.28.0 do not apply any audio decode-size or duration limit when…
vLLM
More for vLLM RSS feed for vLLM
- Severity
- High, CVSS 7.5
- CVSS vector
- CVSS:3.1/AV:N/AC:L/PR:N/UI:N/S:U/C:N/I:N/A:H
- Record
- CVE-2026-90554
- Affects
- vLLM
- Known exploited
- Not listed in the CISA Known Exploited Vulnerabilities catalog
- Published
- Record first published by the source
- September 12, 2026
- Source last modified
- September 24, 2026
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.
Description quoted from the NVD record.
Who is affected
vLLM is software that runs AI language models on a server so that other apps and users can send them requests. This flaw affects vLLM versions from 0.10.2 up to, but not including, 0.28.0. It matters only when the server runs a NanoNemotronVL model with the option to use the audio inside videos turned on. In that setup, an attacker can send a small, highly compressed video that makes the server use gigabytes of memory while pulling out its sound, which can stop the service from working. This flaw is rated High.
What to do
- Find out who installed or runs vLLM for your organization: a staff member, or an IT provider or vendor that manages it for you.
- If someone in your organization runs it, ask them to update vLLM to version
0.28.0or later, which fixes this flaw. - If an IT provider or vendor runs it for you, send them this advisory and ask when the update to
0.28.0will be done. - Until the update is done, ask whoever runs vLLM to check whether any of your servers runs a NanoNemotronVL model with audio in videos turned on, because only those servers are exposed to this flaw.
Analysis written from the source record as of September 24, 2026 and reviewed before publishing.