vllm versions prior to 0.29.0 fail to enforce the VLLM_MAX_AUDIO_CLIP_FILESIZE_MB limit during multimodal chat audio decoding. This allows an unauthenticated attacker to submit audio files that exceed the intended size, causing the library to consume large amounts of memory and CPU while decoding. The result is excessive resource consumption that can degrade performance or lead to service unavailability for legitimate users.

The affected product is vllm by the vllm-project. All releases of vllm before 0.29.0 lack enforcement of the audio clip file size limit and are vulnerable. The specific version range is not explicitly enumerated in the CNA data, but the vulnerability impacts every build prior to 0.29.0.

The CVSS base score of 6.9 indicates a moderate severity. The EPSS score is not available, and the vulnerability is not listed in the CISA KEV catalog. Attackers can exploit this weakness remotely by sending oversized audio files through the public chat endpoints without any authentication. This exploitation path leads to unchecked memory and CPU consumption, posing a moderate but potentially significant risk to service availability.