CVE-2026-115913
vLLM: Phi-4 multimodal audio encoder materializes and retains full attention matrices: a single max-length audio OOM-crashes the vLLM engine
- Medium 6.5
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.
Key facts
- Published
- 26 Oct 2026
- Last changed
- 2 Nov 2026
- Assigned by
- GitHub_M
- EPSS
- 0.329% likely to be exploited in the next 30 days, higher than 26.461% of CVEs (scores for 17 Nov 2026)
- SSVC (CISA)
- Exploitation: none; Automatable: no; Technical impact: partial
Severity scores
Each score is shown as its source gave it. They are not averaged or merged.
- 6.5CVSS 3.1, Mediumfrom the CNA (GitHub_M)
CVSS:3.1/AV:N/AC:L/PR:L/UI:N/S:U/C:N/I:N/A:H
Historical context and remediation
This reference retains the original CVE publication date. A recent record update does not mean a new attack occurred.
No separately sourced impact summary is available on this page yet. Consult the description, version data and dated exploitation evidence.
No explicit remediation statement is available on this page yet. Consult the original record and vendor references before choosing a fix.
Version statuses below come from the CNA. An “unaffected” version is not necessarily a currently supported release or proof of a patch date. EPSS is a dated risk estimate, not a count of historical attacks.
Affected products
| Vendor | Product | Versions |
|---|---|---|
| vllm-project | vllm |
|
Timeline
- 26 Oct 2026 Published