CVE-2026-90554: vLLM before 0.28.0 Denial of Service via audio extraction
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 nanonemotronvl.py, extractaudiofromvideos calls loadaudiopyav(BytesIO(videobytes)) without the maxdurations or maxdecodebytes parameters, so neither VLLMMAXAUDIODECODEDURATIONS nor VLLMMAXAUDIODECODEBYTES is enforced (unlike the direct audio upload path in AudioMediaIO). When a NanoNemotronVL model is served with useaudioinvideo=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 Software
Remediation
Recommended actions to resolve this vulnerability, in priority order.
- Upgrade
Upgrade
vLLMto a version that resolves this vulnerability.Fixed in 0.28.0 - Configuration
For NanoNemotronVL with use_audio_in_video=True, ensure _extract_audio_from_videos passes the max_duration_s and max_decode_bytes parameters to load_audio_pyav(BytesIO(video_bytes)) so VLLM_MAX_AUDIO_DECODE_DURATION_S and VLLM_MAX_AUDIO_DECODE_BYTES are enforced during audio decoding.
vLLM (nano_nemotron_vl.py / _extract_audio_from_videos) Enforce audio decode limits for load_audio_pyav(BytesIO(video_bytes) = Apply VLLM_MAX_AUDIO_DECODE_DURATION_S and VLLM_MAX_AUDIO_DECODE_BYTES when calling load_audio_pyav for video-derived audio extraction
Event History
Frequently Asked Questions
Which deployments are exposed to this issue?
Deployments running vLLM versions 0.10.2 through before 0.28.0 are exposed only when serving a NanoNemotronVL model with use_audio_in_video=True. The affected path is video input processing that extracts audio.
What does an attacker need to do to trigger the denial of service?
An attacker needs to submit a multimodal video input containing highly compressed audio that expands substantially during decoding. No privileges or user interaction are required according to the supplied severity vector.
Do the audio decoding environment limits mitigate this issue?
No. VLLM_MAX_AUDIO_DECODE_DURATION_S and VLLM_MAX_AUDIO_DECODE_BYTES are not enforced on the vulnerable audio-from-video extraction path, even though they are used for direct audio uploads.
What should teams do if they cannot immediately upgrade?
Avoid serving NanoNemotronVL with use_audio_in_video=True or prevent untrusted users from submitting video multimodal inputs. Upgrading to vLLM 0.28.0 fixes the issue.