vLLM audio-transcription path passes IAMF containers to FFmpeg/PyAV, exposing heap-overflow DoS on the API process

vLLM before 0.29.0 passes uploaded audio to the PyAV/FFmpeg av.open() fallback with only a file-size check, exposing FFmpeg's IAMF parser heap overflow. Authenticated remote attackers can repeatedly upload malformed IAMF containers, corrupting the heap and crashing the API process. This is a DoS, not RCE — but it hits the central serving process shared by all concurrent users.

In a shared vLLM serving deployment the API process hosts all model requests and session state. Crashing it on demand with a few small uploads is a reliable availability kill against a production LLM endpoint, and transcriptions may be enabled by default on multimodal serving stacks that accept any authenticated caller.

Authenticated remote attacker POSTs malformed IAMF audio containers to /v1/audio/transcriptions; the endpoint accepts uploaded audio with only a file-size check, and the FFmpeg IAMF parser heap overflow corrupts the heap and crashes the API process.

vLLM < 0.29.0 (PyAV/FFmpeg audio decoding fallback in /v1/audio/transcriptions)

Upgrade to vLLM 0.29.0 (GHSA-m52c-39rh-f3gp). Limit which users can call transcription endpoints and rate-limit uploads until patched.