UpHeron
All vulnerabilities

Servers and applications

CVE-2026-115913

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

Severity
Medium 6.5
Published
20 Oct 2026
Last changed
2 Nov 2026
Added to UpHeron
3 Nov 2026
Fixed in
0.32.0
Also known as
GHSA-8397-8qf6-f2gm
Sources
NVD, CVE Program, Project advisories (GitHub)

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…

Affected software

SoftwareAffected versions
vLLMServers and applications · cna:vllm-project|vllm
  • Every version before 0.32.0
vLLMServers and applications · cpe:a:vllm:vllm
  • From 0.31.0 before 0.32.0
vLLMServers and applications · github:vllm-project/vllm
  • Version 0.31.0

References

  • NVD https://nvd.nist.gov/vuln/detail/CVE-2026-115913
  • CVE.org https://www.cve.org/CVERecord?id=CVE-2026-115913
  • GitHub Advisory Database https://github.com/advisories/GHSA-8397-8qf6-f2gm
  • github.com https://github.com/vllm-project/vllm/security/advisories/GHSA-8397-8qf6-f2gm