The vLLM inference and serving engine for large language models has an uncontrolled resource consumption vulnerability (CWE-400) in versions 0.22.0 through 0.23.0. Specifically, the /v1/audio/transcriptions and /v1/audio/translations endpoints call request.file.read() to fully load uploaded audio files into memory before enforcing the documented maximum compressed upload size limit (VLLM_MAX_AUDIO_CLIP_FILESIZE_MB, default 25 MB). This allows an attacker with access to these routes to submit oversized multipart uploads that cause vLLM to allocate memory proportional to the file size before rejecting the request, resulting in memory pressure or process termination depending on deployment resource limits. The vulnerability is addressed in vLLM version 0.24.0.

An attacker able to access the affected API routes can cause excessive memory consumption by uploading large audio files, leading to denial of service through memory exhaustion or process termination. There is no impact on confidentiality or integrity, only availability is affected.

Upgrade to vLLM version 0.24.0 or later, where this issue is fixed. Until then, restrict access to the affected API routes or implement external upload size checks to prevent oversized uploads from reaching the service. Patch status is confirmed fixed in 0.24.0.