What happened
The vLLM AudioMediaIO multimodal path calls load_audio without the max_duration_s decoded-duration limit, so an authenticated attacker can send a small compressed audio payload via input_audio or audio_url in /v1/chat/completions that decompresses to an oversized decoded duration, exhausting memory/CPU.
Why it matters
A compact request causes disproportionate resource consumption in the multimodal path, enabling a request-driven denial of service against audio-capable vLLM serving instances (CVSS 6.5, CWE-409).
Attack vector
Send a small, highly-compressed audio payload via input_audio/audio_url to trigger decompression-bomb resource exhaustion.
Affected systems
vLLM >= 0.19.0, < 0.24.0 (AudioMediaIO path).
Mitigation
Upgrade to vLLM 0.24.0; the fix applies the max_duration_s decoded-duration limit on the load_audio path.