---
id: CVE-2026-115913
title: >-
  vLLM: Phi-4 multimodal audio encoder materializes and retains full attention
  matrices: a single max-length audio OOM-crashes the vLLM engine
summary: >-
  vLLM is an inference and serving engine for large language models. Prior to
  0.32.0, the Phi-4 multimodal audio encoder in phi4mm_audio.py and
  phi4mm_utils.py uses eager attention by default, allowing a remote client to
  submit one legal l…
severity: medium
cvss: 6.5
cvssVector: 'CVSS:3.1/AV:N/AC:L/PR:L/UI:N/S:U/C:N/I:N/A:H'
cvssSource: cna
cwe:
  - CWE-770
vendor: vllm-project
product: vllm
affected:
  - vllm < 0.32.0
ssvc:
  exploitation: none
  automatable: 'no'
  technicalImpact: partial
  timestamp: '2026-11-02T17:48:03.108418Z'
published: '2026-10-26'
updated: '2026-11-02'
sourceUpdated: '2026-11-02T17:48:14.216Z'
source: CVEORG
sourceUrl: 'https://www.cve.org/CVERecord?id=CVE-2026-115913'
references:
  - url: >-
      https://github.com/vllm-project/vllm/security/advisories/GHSA-8397-8qf6-f2gm
    label: >-
      https://github.com/vllm-project/vllm/security/advisories/GHSA-8397-8qf6-f2gm
  - url: 'https://github.com/vllm-project/vllm/pull/60614'
    label: 'https://github.com/vllm-project/vllm/pull/60614'
  - url: >-
      https://github.com/vllm-project/vllm/commit/3d52578e13454de0bb8352cac988347cdcc89928
    label: >-
      https://github.com/vllm-project/vllm/commit/3d52578e13454de0bb8352cac988347cdcc89928
  - url: 'https://github.com/vllm-project/vllm/releases/tag/v0.32.0'
    label: 'https://github.com/vllm-project/vllm/releases/tag/v0.32.0'
  - url: 'https://nvd.nist.gov/vuln/detail/CVE-2026-115913'
  - url: 'https://github.com/vllm-project/vllm'
  - url: 'https://github.com/advisories/GHSA-8397-8qf6-f2gm'
tags:
  - cve.org
  - osv
  - pip
  - ghsa
epss: 0.00329
epssPercentile: 0.26221
aliases:
  - GHSA-8397-8qf6-f2gm
ecosystem: pip
patched:
  - vllm 0.32.0
ingestedAt: '2026-10-26T17:33:33.061Z'
---

## Overview

vLLM is an inference and serving engine for large language models. Prior to 0.32.0, the Phi-4 multimodal audio encoder in phi4mm_audio.py and phi4mm_utils.py uses eager attention by default, allowing a remote client to submit one legal long audio clip that causes MultiHeadedAttention.forward_attention to materialize and retain full attention matrices in self.attn. The scaled dot-product attention branch does not materialize these matrices. The resulting quadratic GPU-memory use can cause an out-of-memory condition and terminate EngineCore for all users. This issue is fixed in version 0.32.0.

## Affected

- `vllm < 0.32.0`

## Remediation

Refer to the linked advisories for vendor-supplied fixes and affected version ranges.

## Package advisory (CVE-2026-115913)

Affected packages:

- `vllm < 0.32.0`

Patched in:

- `vllm 0.32.0`

Source: https://osv.dev/vulnerability/GHSA-8397-8qf6-f2gm