Executive brief
vLLM is a high-throughput engine for serving AI models. A vulnerability in its speech-to-text interface allows an authorized user to upload excessively large audio files that bypass intended size limits. This can cause the server to run out of memory and crash, leading to a service outage for all users.
Technical details
The vLLM speech-to-text API endpoints (/v1/audio/transcriptions and /v1/audio/translations) perform an unbounded read of uploaded files using FastAPI's UploadFile.read() before validating the file size against the VLLM_MAX_AUDIO_CLIP_FILESIZE_MB configuration. This root cause allows an attacker with API access to submit oversized multipart uploads that are materialized into memory as bytes objects, potentially leading to process termination due to memory pressure (CWE-770/CWE-400). The vulnerability is present in the api_router.py components for transcription and translation. A fix is available in version 0.24.0 which implements bounded reads and early size metadata checks.
Affected products
- vllm-project vllm >= 0.22.0, < 0.24.0
Timeline
- 2026-07-02: disclosed
- 2026-07-06: advisory: NVD publication date
- 2026-07-17: patched: GitHub Advisory reviewed and updated with patch information