Vllm · Vllm · CVE-2026-54233
**Name of the Vulnerable Software and Affected Versions**
vLLM versions prior to 0.23.1rc0
**Description**
The '/v1/audio/transcriptions' endpoint limits the size of compressed uploads but fails to limit the size of the decoded PCM (Pulse Code Modulation) output. PCM is an uncompressed digital audio format. The `SpeechToTextProcessor` rejects uploads exceeding the `VLLM MAX AUDIO CLIP FILESIZE MB` limit based on compressed byte length; however, the audio decoder in `audio.py` accumulates all decoded frames into memory without a size limit. For example, a 25MB OPUS file can expand to approximately 14.9GB of float32 PCM during decode time because the `np.concatenate` function allocates a second contiguous array. The `SpeechToTextConfig.max audio clip s` limit is only applied after the full decode process, which does not prevent the initial memory allocation. This allows an unauthenticated attacker to exhaust server memory using a small number of concurrent requests that stay within the documented upload size limits.
**Recommendations**
Update to version 0.23.1rc0 or later.