Back to CVE records
Medium 6.5Published Oct 26, 2026

CVE-2026-115913vLLM: Phi-4 multimodal audio encoder materializes and retains full attention matrices: a single max-length audio OOM-crashes the vLLM engine

Updated under an hour ago ()

Does this affect you?

It affects vllm-project vllm and 1 other product. Add what you run to find out whether that includes you.

No exploitation reported and no EPSS estimate yet.

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.

Free: a daily email for up to 3 products and the weekly digest. No card, no password. By continuing you agree to the Terms and Privacy Policy.

CVSS
6.5 MEDIUM (v3.1, CNA)CVSS:3.1/AV:N/AC:L/PR:L/UI:N/S:U/C:N/I:N/A:H
Weakness (CWE)
CWE-770
Known exploited
Not in CISA's KEV catalog
EPSS (exploitation in the next 30 days)
Not loaded for this CVE
Assigned by
GitHub_M
NVD analysis
Analyzed
Record last updated
Oct 26, 2026
First seen by InfoSec Signals
Oct 26, 2026

Affected products

Affected products
vllm-project vllm
vllm_project:vllm (from the name)
before 0.32.0
CNA record
vllm vllm
vllm:vllm
from 0.31.0 before 0.32.0
NVD

References

References

The CVE record is authoritative. Products marked “from the name” were identified from the CNA's free-text names, not a CPE, and are matched less certainly. This product uses the NVD API but is not endorsed or certified by the NVD.

Essential cookies only, unless you allow a Google ads cookie. See our .