CVE-2026-57173: vLLM: Unauthenticated audio decompression-bomb DoS in /v1/chat/completions

vLLM has a decode guard, max_duration_s, set from VLLM_MAX_AUDIO_DECODE_DURATION_S with a default of 600 seconds. It rejects long audio during decode, before memory is allocated. The code passes it in only one place: the speech-to-text serving layer for /v1/audio/transcriptions.

The chat path never passes it. An input_audio content part goes through parse_input_audio, connector.fetch_audio, AudioMediaIO._load_data_url, load_base64 and load_bytes, which calls load_audio(BytesIO(data), sr=None). With no limit set, every length check in load_audio_pyav and load_audio_soundfile is skipped. Inline data: URLs need no HTTP fetch, so VLLM_AUDIO_FETCH_TIMEOUT does not help either.

The attacker sends one small, highly compressed audio file as a base64 input_audio part. It decodes into float32 samples that take several GB, and the worker is OOM-killed. The OpenAI-compatible server has no authentication unless --api-key or VLLM_API_KEY is set, so the advisory calls it unauthenticated. The CVSS vector lists low privileges. Any deployment serving an audio-capable model is affected. The advisory names two sibling bugs in the same media code, for video frames and for images.

Affected: vllm 0.23.0 and earlier, on servers that run an audio-capable model behind /v1/chat/completions. Servers without such a model do not decode audio. Check the version:

pip show vllm

The speech-to-text endpoint already had the guard.

Upgrade to vLLM 0.24.0.

pip install --upgrade "vllm>=0.24.0"

The fix is pull request 45908. The sources list no workaround. Until you upgrade, set an API key with --api-key or VLLM_API_KEY and keep the server off untrusted networks, or limit request body size in a proxy in front of it.

Look for worker or engine processes killed by the out-of-memory killer, or for container restarts, right after a chat completion request that contained an input_audio part. In request logs, look for small bodies with base64 audio from a single client that precede memory spikes. A body of a few KB with a memory jump of GBs points to the attack.

Summary

The audio decode-duration guard (max_duration_s, env VLLM_MAX_AUDIO_DECODE_DURATION_S, default 600s) that protects against audio decompression-bomb DoS is wired into only the speech-to-text path (/v1/audio/transcriptions). The chat audio path (/v1/chat/completions, input_audio content parts) calls the same decoder with no limit, so an unauthenticated client can submit a few-KB compressed audio file that expands to multiple GB of float32 PCM at decode time, OOM-killing the worker. This is a distinct sibling of CVE-2026-5497 (video frame-count bomb, VideoMediaIO.load_base64) and GHSA-pq5c-rjhq-qp7p (image) in the same media subsystem.

Verified against main at HEAD d78650c (2026-06-16); applicable to the latest release v0.23.0.

Details

The guard rejects long audio *during* decode (before allocation), implemented in vllm/multimodal/media/audio.py:

• load_audio_pyav — metadata reject (~82-98) and live sample-count reject (~129-136)
• load_audio_soundfile — frames reject (~165-174)

All are gated on if max_duration_s is not None.

It is passed in exactly one place — the transcription serving layer:

# .../speech_to_text/base/serving.py:~170-174
load_audio(buf, sr=..., max_duration_s=self.max_audio_decode_duration_s)
#   self.max_audio_decode_duration_s = envs.VLLM_MAX_AUDIO_DECODE_DURATION_S  (default 600)

The chat path never threads it:

# vllm/multimodal/media/audio.py:237-238
def load_bytes(self, data: bytes) -> tuple[npt.NDArray, float]:
    return load_audio(BytesIO(data), sr=None)   # no max_duration_s -> every guard above is skipped

Unauthenticated reachability chain (chat): parse_input_audio (chat_utils.py) -> parse_audio -> connector.fetch_audio -> AudioMediaIO._load_data_url -> load_base64 -> load_bytes -> load_audio(..., sr=None). The connector never passes max_duration_s, and inline data: URLs need no HTTP fetch (so VLLM_AUDIO_FETCH_TIMEOUT does not bound them). The OpenAI-compatible server has no auth by default (auth only when --api-key / VLLM_API_KEY is set).

Impact

Unauthenticated remote denial of service (availability) via memory amplification on a default-no-auth endpoint, on any deployment serving an audio-capable model. Same class and impact as the sibling CVE-2026-5497 (video). CWE-770 / CWE-409.

Fix

A fix was introduced in this MR: https://github.com/vllm-project/vllm/pull/45908