vLLM: Phi-4 multimodal audio encoder materializes and retains full attention matrices: a single max-length audio OOM-crashes the vLLM engine
Summary
| CVE | CVE-2026-115913 |
|---|---|
| State | PUBLISHED |
| Assigner | GitHub_M |
| Source Priority | CVE Program / NVD first with legacy fallback |
| Published | 2026-10-26 17:17:19 UTC |
| Updated | 2026-10-28 15:38:12 UTC |
| 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. |
Risk And Classification
Primary CVSS: v3.1 6.5 MEDIUM from [email protected]
CVSS:3.1/AV:N/AC:L/PR:L/UI:N/S:U/C:N/I:N/A:H
EPSS: 0.003050000 probability, percentile 0.232050000 (date 2026-10-28)
Problem Types: CWE-770 | CWE-770 CWE-770: Allocation of Resources Without Limits or Throttling
| Version | Source | Type | Score | Severity | Vector |
|---|---|---|---|---|---|
| 3.1 | [email protected] | Secondary | 6.5 | MEDIUM | CVSS:3.1/AV:N/AC:L/PR:L/UI:N/S:U/C:N/I:N/A:H |
| 3.1 | CNA | DECLARED | 6.5 | MEDIUM | CVSS:3.1/AV:N/AC:L/PR:L/UI:N/S:U/C:N/I:N/A:H |
CVSS v3.1 Breakdown
Attack Vector
NetworkAttack Complexity
LowPrivileges Required
LowUser Interaction
NoneScope
UnchangedConfidentiality
NoneIntegrity
NoneAvailability
HighCVSS:3.1/AV:N/AC:L/PR:L/UI:N/S:U/C:N/I:N/A:H
NVD Known Affected Configurations (CPE 2.3)
Vendor Declared Affected Products
| Source | Vendor | Product | Version | Platforms |
|---|---|---|---|---|
| CNA | Vllm-project | Vllm | affected < 0.32.0 | Not specified |
References
| Reference | Source | Link | Tags |
|---|---|---|---|
| github.com/vllm-project/vllm/releases/tag/v0.32.0 | [email protected] | github.com | Release Notes |
| github.com/vllm-project/vllm/security/advisories/GHSA-8397-8qf6-f2gm | [email protected] | github.com | Exploit, Patch, Vendor Advisory |
| github.com/vllm-project/vllm/commit/3d52578e13454de0bb8352cac988347cdcc8... | [email protected] | github.com | Patch |
| github.com/vllm-project/vllm/pull/60614 | [email protected] | github.com | Patch |
| CVE Program record | CVE.ORG | www.cve.org | canonical |
| NVD vulnerability detail | NVD | nvd.nist.gov | canonical, analysis |
No vendor comments have been submitted for this CVE.
There are currently no legacy QID mappings associated with this CVE.