Technical impact
vLLM versions from 0.19.0 up to before 0.24.0 contain a decompression bomb vulnerability in the AudioMediaIO multimodal path. Maliciously sized audio files sent via the input_audio or audio_url parameter in the /v1/chat/completions endpoint are decoded into an unbounded PCM buffer. The inflated buffer consumes large amounts of memory, eventually crashing or killing the server process, resulting in a denial of service for all users interacting with the model. This flaw corresponds to CWE‑409, an inappropriate restriction of operations that can cause resource exhaustion.
Affected Systems
All releases of vLLM from 0.19.0 through 0.23.x are affected. The vulnerable product is the official vLLM library from the vllm‑project. The issue is reported to have been resolved in v0.24.0 and later. The known platform enumeration string matches the main vLLM package.
Risk and Exploitability
The CVSS score of 7.1 indicates moderate‑severe risk. EPSS data is not available, so the exploitation probability is currently unknown. The vulnerability does not appear in the CISA KEV catalog. Because authentication to the chat endpoint is required, attackers need valid credentials or a session with the model. Once authenticated, an attacker can submit a small compressed audio resource that expands to a large PCM buffer, exhausting memory and triggering a crash or slowdown. The impact is limited to service availability rather than confidentiality or integrity.