vLLM versions >=0.10.2 and <0.28.0 do not enforce audio decode-size or duration limits when extracting audio from video input for NanoNemotronVL models. Specifically, in nano_nemotron_vl.py, the function _extract_audio_from_videos calls load_audio_pyav(BytesIO(video_bytes)) without passing max_duration_s or max_decode_bytes parameters, so the environment variables VLLM_MAX_AUDIO_DECODE_DURATION_S and VLLM_MAX_AUDIO_DECODE_BYTES are not applied. When use_audio_in_video=True, an attacker can supply a small, highly compressed video that forces the server to allocate gigabytes of memory during audio decoding, resulting in denial of service. This issue was fixed in vLLM version 0.28.0.

An attacker can cause a denial of service by supplying specially crafted video input that triggers excessive memory allocation during audio decoding. This can exhaust server resources and disrupt service availability. The vulnerability does not require privileges or user interaction and has a medium severity score (CVSS 6.9).

Upgrade to vLLM version 0.28.0 or later, where this vulnerability is fixed. No other mitigation is indicated or required.