The vulnerability arises because vLLM versions between 0.10.2 and before 0.28.0 allow unlimited audio extraction when a video is supplied to NanoNemotronVL models configured with use_audio_in_video=True. The extraction routine does not enforce the configuration limits VLLM_MAX_AUDIO_DECODE_DURATION_S or VLLM_MAX_AUDIO_DECODE_BYTES, causing the server to allocate memory proportional to the audio stream size. An attacker can craft a small, highly compressed video that forces the server to decode gigabytes of audio, exhausting memory and causing a service halt. This represents a resource exhaustion flaw identified as CWE-400.

vllm-project's vLLM product is impacted when the package version is greater than or equal to 0.10.2 and less than 0.28.0. The failure affects deployments that expose NanoNemotronVL models with audio extraction enabled. No other vendors or products are listed.

The CVSS score of 6.9 indicates a moderate severity. The EPSS score is not available, so the prevalence of vulnerability is not listed in the CISA KEV catalog. The likely attack vector, inferred from the description, is an adversary sending malicious multimodal input via the server’s inference endpoint; it requires only the ability to submit video payloads and no privileged access. If triggered, the entire service can become unavailable due to memory exhaustion.