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.
Published: 2026-10-26
Score: 6.5 Medium
EPSS: n/a
KEV: No
Impact: Denial of Service (CWE-770)
Action: Immediate Patch
AI Analysis

Impact

The vulnerability originates from vLLM's Phi‑4 multimodal audio encoder in versions before 0.32.0. By using eager attention, the engine materializes full attention matrices. When a single long audio clip is processed, the quadratic GPU‑memory usage can cause an out‑of‑memory crash that terminates EngineCore and renders the inference service unavailable to all users. This is a Resource Exhaustion defect (CWE‑770).

Affected Systems

Any deployment of vllm-project:vllm using a pre‑0.32.0 build is affected. The issue is fixed in v0.32.0 and subsequent releases.

Risk and Exploitability

The CVSS score of 6.5 indicates moderate severity. Although EPSS is not reported, a remote client can trigger the vulnerability by sending a sufficiently long legal audio clip to the inference endpoint. The exploit requires only network access and can result in a denial of service for all users. The vulnerability is not listed in the CISA KEV catalog.

Generated by OpenCVE AI on October 26, 2026 at 18:52 UTC.

Remediation

No solution or workaround provided in the CVE record.

OpenCVE Recommended Actions

  • Upgrade vLLM to v0.32.0 or newer
  • Restart the vLLM engine after the update to load the new code
  • If upgrading immediately is not possible, limit the maximum length of submitted audio clips or configure the engine to disable eager attention to prevent large memory allocations

Generated by OpenCVE AI on October 26, 2026 at 18:52 UTC.

Tracking

Sign in to view the affected projects.

Advisories

No advisories yet.

History

Mon, 26 Oct 2026 18:45:00 +0000

Type Values Removed Values Added
First Time appeared Vllm
Vllm vllm
Vllm-project
Vllm-project vllm
Vendors & Products Vllm
Vllm vllm
Vllm-project
Vllm-project vllm

Mon, 26 Oct 2026 17:15:00 +0000

Type Values Removed Values Added
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.
Title vLLM: Phi-4 multimodal audio encoder materializes and retains full attention matrices: a single max-length audio OOM-crashes the vLLM engine
Weaknesses CWE-770
References
Metrics cvssV3_1

{'score': 6.5, 'vector': 'CVSS:3.1/AV:N/AC:L/PR:L/UI:N/S:U/C:N/I:N/A:H'}

Subscriptions

Vllm Vllm
Vllm-project Vllm
cve-icon MITRE

Status: PUBLISHED

Assigner: GitHub_M

Published:

Updated: 2026-10-26T16:56:59.817Z

Reserved: 2026-10-23T20:21:45.107Z

Link: CVE-2026-115913

cve-icon Vulnrichment

No data.

cve-icon NVD

Status : Awaiting Analysis

Published: 2026-10-26T17:17:19.053

Modified: 2026-10-26T18:34:17.050

Link: CVE-2026-115913

cve-icon Redhat

No data.

cve-icon OpenCVE Enrichment

Updated: 2026-10-26T19:00:08Z

Weaknesses
  • CWE-770

    Allocation of Resources Without Limits or Throttling