{"id":"CVE-2026-54233","aliases":["GHSA-6pr9-rp53-2pmc","PYSEC-2026-3404"],"title":"vLLM: OOM Denial of Service via Audio Decompression Bomb","summary":"vLLM: OOM Denial of Service via Audio Decompression Bomb","severity":"medium","cvss":6.5,"cvssVector":"CVSS:3.1/AV:N/AC:L/PR:L/UI:N/S:U/C:N/I:N/A:H","vendor":"vllm","product":"vllm","ecosystem":"pip","affected":["vllm < 0.24.0"],"patched":["vllm 0.24.0"],"published":"2026-06-17","updated":"2026-09-10","sourceUpdated":"2026-09-10T03:51:08.015014698Z","source":"OSV","sourceUrl":"https://osv.dev/vulnerability/GHSA-6pr9-rp53-2pmc","references":[{"url":"https://github.com/vllm-project/vllm/security/advisories/GHSA-6pr9-rp53-2pmc"},{"url":"https://nvd.nist.gov/vuln/detail/CVE-2026-54233"},{"url":"https://github.com/vllm-project/vllm/pull/44970"},{"url":"https://github.com/vllm-project/vllm/commit/1b1359c33269446f13c05da9a90c25174cbea590"},{"url":"https://github.com/advisories/GHSA-6pr9-rp53-2pmc"},{"url":"https://github.com/pypa/advisory-database/tree/main/vulns/vllm/PYSEC-2026-3404.yaml"},{"url":"https://github.com/vllm-project/vllm"},{"url":"https://github.com/vllm-project/vllm/releases/tag/v0.23.1rc0"},{"url":"https://pypi.org/project/vllm"}],"tags":["osv","pip","ghsa"],"epss":0.00422,"epssPercentile":0.36014,"cwe":["CWE-409"],"ingestedAt":"2026-06-29T14:31:47.437Z","slug":"CVE-2026-54233","body":"## Overview\n\n### Summary\nvLLM's `/v1/audio/transcriptions` endpoint limits compressed upload size but not decoded PCM output. A 25MB OPUS file expands to ~14.9GB of float32 PCM at decode time. Tested on vLLM v0.19.0.\n\n### Details\n`SpeechToTextProcessor` rejects uploads over `VLLM_MAX_AUDIO_CLIP_FILESIZE_MB` (default 25MB) based on compressed byte length, but the audio decoder in `audio.py` accumulates all decoded frames into memory with no size limit before returning:\n\n```python\n# speech_to_text.py L184-189\nif len(audio_data) / 1024 ** 2 > self.max_audio_filesize_mb:\n    raise VLLMValidationError(...)\ny, sr = load_audio(buf, sr=self.asr_config.sample_rate)  # decoded size unchecked\n\n# audio.py L77-107\nchunks: list[npt.NDArray] = []\nfor frame in container.decode(stream):\n    chunks.append(frame.to_ndarray())\naudio = np.concatenate(chunks, axis=-1).astype(np.float32)  # single contiguous allocation\n```\n\nA 25MB OPUS file at 6kbps encodes ~8.7 hours of audio. Decoding produces ~5.7GB of float32 PCM (232x amplification), and `np.concatenate` then allocates a second contiguous array, bringing peak RSS to ~14.9GB from a single request. `SpeechToTextConfig.max_audio_clip_s` (default 30s) applies only after the full decode and does not prevent the allocation.\n\n### Impact\nAn unauthenticated attacker can exhaust server memory with a small number of concurrent requests, each a valid upload within the documented size limit. Severity was assessed with reference to prior OOM vulnerability reports in vLLM.\n\n### Fix\n\nA fix for this vulnerability was merged here: https://github.com/vllm-project/vllm/pull/44970\n\n## Affected packages\n\n- `vllm < 0.24.0`\n\n## Remediation\n\nUpgrade to a patched release:\n\n- `vllm 0.24.0`","depth":"sunlit","depthScore":36,"depthScoreParts":{"impact":35.8,"likelihood":0.1,"exploitation":0,"ransomware":0},"changes":[]}