Skip to content
Fencer Vulns
medium 6.5

CVE-2026-115913 json

vLLM: Phi-4 multimodal audio encoder materializes and retains full attention matrices: a single max-length audio OOM-crashes the vLLM engine

published 2026-10-26 13 d ago modified 2026-11-02 6 d ago also GHSA-8397-8qf6-f2gm , EUVD-2026-110023

cvss
6.5
version 3.1, CNA
cisa kev
Not listed
not in CISA's catalog
epss
0.3%
EPSS 0.3% chance of exploitation in 30 days, percentile 26%
public exploits
0
none known

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 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.

Patch / fix

  • pypi vllm 0.32.0
  • vllm vllm 0.32.0

CISA SSVC

Exploitation
none
Automatable
no
Technical impact
partial
Assessed
6 d ago

Exploit prediction

0.33% chance of exploitation in the next 30 days, higher than 26% of all scored CVEs (FIRST EPSS, 2026-11-08).

Exploits and checks

  • github.com https://github.com/vllm-project/vllm/security/advisories/GHSA-8397-8qf6-f2gm referenced PoC, tagged "exploit" by a source, unverified

CVSS scores

assigned by version score vector
CNA 3.1 6.5 CVSS:3.1/AV:N/AC:L/PR:L/UI:N/S:U/C:N/I:N/A:H
GitHub 3.1 6.5 CVSS:3.1/AV:N/AC:L/PR:L/UI:N/S:U/C:N/I:N/A:H

Affected packages

ecosystem package affected fixed in
pypi vllm
<0.32.0
0.32.0

Affected products

  • vllm-project vllm

Classification

Weakness
  • CWE-770 Allocation of Resources Without Limits or Throttling

References

Sources

Fencer Vulns keeps one entry per vulnerability: the CVE id wins, then GHSA, and records that alias each other are merged, except that two CVEs always keep separate entries. The same data is available as JSON.