CVE-2026-115913

Description

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.

Severity
MEDIUM
CVSS v3
6.5
CVSS:3.1/AV:N/AC:L/PR:L/UI:N/S:U/C:N/I:N/A:H · source cna
Exploit chance (EPSS)
0.329%
chance of exploitation in the next 30 days
higher than 26.5% of scored CVEs · 2026-11-17
CISA KEV
Not listed
Public exploit code
Yes
Exploit references (NVD and CNA): 1
Metasploit modules: 0
Nuclei templates: 0

Weakness

CISA SSVC decision points

Exploitationnone
Automatableno
Technical impactpartial
Scored2026-11-02

Timeline

  1. 2026-10-26Published by the CNA16:56 UTC
  2. 2026-10-28GitHub advisory published (GHSA-8397-8qf6-f2gm)2 days after publication
  3. 2026-11-02CISA SSVC decision recorded7 days after publication

The timeline shows what already happened. To be told when this CVE changes next, watch it from the API.

Exploit and advisory records

GitHub advisory GHSA-8397-8qf6-f2gm, affected packages:

PackageVulnerable versionsFixed in
PyPI vllmbefore 0.32.00.32.0

Affected products

Products listed as vulnerable. Upgrade to the version under "Fixed in", or a later one.

ProductVulnerable versionsFixed in
vllm vllmfrom 0.31.0 before 0.32.0fixed in 0.32.00.32.0

Products and versions as NVD and the vendor that published the CVE list them.

References

Get more CVE data from the API.

Free without a key or signup. Pro adds monitoring and higher limits: pricing.

Sources