CVE Professor
›
Back
CVE-2026-115913

CVE-2026-115913 — Vllm Vllm (tüm Sürümler) — CWE-770

MEDIUM CVSS 6.5 NVD

CVE-2026-115913 is a MEDIUM severity security vulnerability (CWE-770) (CVSS 6.5/10) affecting Vllm Vllm (tüm Sürümler). Technical vector analysis shows that an attacker can exploit this flaw remotely over the network with low-privileged access (PR:L) with zero user interaction (UI:N), causing high-impact service disruption (A:H).

📄 Official NVD / MITRE Original Record Text

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 26.10.2026
ISO Date 2026-10-26
CVSS Version v3.1
Source NVD
CVSS v3.1 Vector Metrics
CVSS:3.1/AV:N/AC:L/PR:L/UI:N/S:U/C:N/I:N/A:H
Attack Vector Network AV:N
Attack Complexity Low AC:L
Privileges Req. Low PR:L
User Interaction None UI:N
Scope Unchanged S:U
Confidentiality None C:N
Integrity None I:N
Availability High A:H
AI Security Analysis & Remediation
🤖 Generating Deep AI Security Analysis…
Analyzing CVSS vector & CWE weakness structure…

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
CVE-2026-115913
CVSS 6.5/10 · MEDIUM · 26.10.2026
CWE: CWE-770
Vector: CVSS:3.1/AV:N/AC:L/PR:L/UI:N/S:U/C:N/I:N/A:H
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

🎯 WHAT IS THIS VULNERABILITY?
Summary: 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 attenti…
Affected: vllm:vllm

⚡ ATTACK CHAIN
Remotely over internet | low-privilege account | no user interaction | low complexity.

🌍 REAL WORLD IMPACT
Service can be completely taken down.

🛡️ REMEDIATION
→ Apply vendor security update for vllm:vllm.
→ Add WAF rule, create IDS alert for anomalous traffic.

⚠️ URGENCY
📋 Routine update cycle (30 days)

────────────────────────────────────────
[Kural Tabanlı Otomatik Güvenlik Analizi]

Target Reconnaissance Queries (FOFA & Shodan)

FOFA Query

app="vllm vllm"
CVSS Score
6.5 / 10 · CVSS MEDIUM
Weakness Types (CWE)
CWE-770
Affected Products (1)
vllm:vllm (tüm sürümler)
References & Advisories (4)
github.com patch
github.com patch
github.com patch exploit advisory
Related CVEs
6.5
CVE-2026-132750
Meilisearch before 1.55.1 and 1.56.x before 1.56.1 contains a quadratic memory consumption vulnerability in fi
8.7
CVE-2026-130761
Fleet before 4.92.0 contains a denial of service vulnerability that allows enrolled hosts to trigger unbounded
8.7
CVE-2026-130749
Grav from 2.0.0 before 2.2.18 contains a resource exhaustion vulnerability that allows unauthenticated attacke
8.7
CVE-2026-130092
A caller who can reach the temporal-proxy gRPC listener, with or without a valid credential, can grow the prox
6.5
CVE-2026-121591
Mattermost Desktop App versions <=6.3 6.0.3 6.3.2.0 fail to bound the Calls widget resize dimensions before ap
6.5
CVE-2026-16021
Mattermost versions 11.11.x <= 11.11.0, 11.7.x <= 11.7.11, 11.10.x <= 11.10.2 fail to bound the resources cons