---
id: "CVE-2026-116175"
kind: "vulnerability"
aliases: ["CVE-2026-116175"]
title: "vLLM: small compressed audio expands into memory-exhausting decoded PCM"
severity: "high"
cvss: 7.1
epss: null
exploited: false
status: "open"
category: "ai-tooling"
products: []
ecosystems: []
tags: ["ai-inference", "ai-tooling", "python", "self-hosted", "vllm"]
not_affected_if: ["auth-required", "internal-only", "text-only-models"]
published_at: "2026-10-25T14:16:57+00:00"
first_seen_at: "2026-10-25T14:41:23+00:00"
updated_at: "2026-10-25T14:42:36+00:00"
sources: [{"source": "nvd", "url": "https://nvd.nist.gov/vuln/detail/CVE-2026-116175", "first_seen_at": "2026-10-25T14:41:23+00:00"}]
---

# vLLM: small compressed audio expands into memory-exhausting decoded PCM

vLLM from 0.19.0 before 0.24.0 passes audio from input_audio or audio_url in chat completion requests to load_audio without the max_duration_s decoded-duration limit. An authenticated attacker can send a small compressed audio file that decodes to a huge PCM stream and exhaust server memory.

_Summary, check and fix written by a model from the sources; labelled as such._

## How to check

- Check the installed version with `pip show vllm` and treat anything from `0.19.0` up to but not including `0.24.0` as exposed.
- Check whether the server accepts audio in chat completions by looking for audio multimodal input enabled on `/v1/chat/completions`.

## Fix

- Upgrade vLLM to `0.24.0` or later.
- Until upgraded, restrict who can reach `/v1/chat/completions` and put a request size or memory limit in front of the server.

## Timeline

- 2026-10-25T14:41:23+00:00 nvd https://nvd.nist.gov/vuln/detail/CVE-2026-116175

## From the sources

vLLM from 0.19.0 before 0.24.0 contains a decompression bomb vulnerability because the AudioMediaIO multimodal path calls load_audio without the max_duration_s decoded-duration limit. Authenticated attackers can send small compressed audio via input_audio or audio_url in /v1/chat/completions requests that expands to huge decoded PCM, exhausting server memory.

---

_Free to read at a human pace. An API with keys and an MCP server are in preparation; early access: hello@infrastack.ai_