Memory Management Issue in vLLM Audio Encoder by VLLM Project
CVE-2026-115913

6.5MEDIUM

Key Information:

Status
Vendor
CVE Published:
26 October 2026

What is CVE-2026-115913?

The vLLM project has identified a memory management issue in its Phi-4 multimodal audio encoder, present in versions prior to 0.32.0. Due to the default use of eager attention in the encoding process, a remote client can exploit this vulnerability by submitting a long audio clip. This action can trigger the MultiHeadedAttention.forward_attention method, leading to the retention of full attention matrices in the self.attn component. As a consequence, this results in excessive GPU memory consumption which may cause an out-of-memory condition, adversely impacting the EngineCore and causing disruptions for all users. This vulnerability has been addressed in version 0.32.0.

Affected Version(s)

vllm < 0.32.0

References

CVSS V3.1

Score:
6.5
Severity:
MEDIUM
Confidentiality:
None
Integrity:
None
Availability:
None
Attack Vector:
Network
Attack Complexity:
Low
Privileges Required:
Low
User Interaction:
None
Scope:
Unchanged

Timeline

  • Vulnerability published

  • Vulnerability Reserved

.