CVE-2026-57173

Name of the Vulnerable Software and Affected Versions vLLM versions prior to 0.24.0

Description An issue exists in the audio handling path for the /v1/chat/completions endpoint where the input audio content is processed without a duration limit. Specifically, the AudioMediaIO.load bytes() and AudioMediaIO.load file() functions call the shared audio decoder without passing the VLLM MAX AUDIO DECODE DURATION S variable. This allows an unauthenticated client to submit a small compressed audio file that expands into a massive float32 PCM allocation during decoding, bypassing the duration guards used in other paths. This memory amplification can lead to an out-of-memory (OOM) worker crash, resulting in a remote denial of service. Additionally, inline data URLs can reach this path without being restricted by the VLLM AUDIO FETCH TIMEOUT setting. This issue affects any deployment serving an audio-capable model.

Recommendations Update vLLM to version 0.24.0. As a temporary mitigation, restrict access to the /v1/chat/completions endpoint or implement authentication to limit reachability.