CyberRota Analysis
AI-GeneratedVersions of vLLM from 0.10.2 to 0.27.9 are vulnerable due to a lack of limits on audio decode size and duration when processing video input for NanoNemotronVL models, allowing attackers to exploit this by submitting small, compressed videos. This can lead to excessive memory allocation during audio decoding, potentially causing a denial of service. Organizations utilizing affected versions of vLLM should prioritize upgrading to version 0.28.0 to mitigate this risk.
Public Exploit Signal
A public exploit, PoC, GitHub repository or Metasploit reference was detected for this CVE.
Note: these links are listed for security research and verification purposes only.
Original NVD 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.