Vulnerability  ·  2026-09-18

vLLM — unbounded audio decode duration in chat-completions input_audio path enables resource-exhaustion DoS

VulnerabilityMedium impactGlobalCVE-2026-57173
NVD published this CVSS 6.5 Medium vulnerability on Sept 16, 2026, part of a recurring pattern of missing resource limits in vLLM's multimodal (audio/video) input-handling paths.
vLLM is one of the most widely deployed open-source LLM inference engines; missing resource caps on multimodal input processing let any client degrade or crash a shared multi-tenant inference server, denying service to all co-hosted users/models.
The input_audio handling path calls AudioMediaIO.load_bytes/load_file without passing the configured VLLM_MAX_AUDIO_DECODE_DURATION_S limit to the shared audio decoder, letting an unauthenticated client submit an oversized audio payload that consumes excessive decode resources on the shared inference server.
vLLM < 0.24.0
Upgrade to vLLM >= 0.24.0, which threads the duration limit through to the audio decoder.
GitHub commit fix
See this in the live feed Explore related AI security and governance findings — updated every morning.
Open the feed →