Vulnerability Details CVE-2026-90554
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.
Exploit prediction scoring system (EPSS) score
EPSS Score 0.002
EPSS Ranking 4.5%
CVSS Severity
CVSS v3 Score 6.2
References
Products affected by CVE-2026-90554
-
cpe:2.3:a:vllm:vllm:0.25.0
-
cpe:2.3:a:vllm:vllm:0.25.1
-
cpe:2.3:a:vllm:vllm:0.26.0
-
cpe:2.3:a:vllm:vllm:0.26.1
-
cpe:2.3:a:vllm:vllm:0.27.0
-
cpe:2.3:a:vllm:vllm:0.27.1