CVE-2026-114055: Uncontrolled Resource Consumption in vLLM via Unvalidated Multimodal Kwargs

Vulnerability ID: CVE-2026-114055
CVSS Score: 7.5
Published: 2026-10-23

vLLM prior to version 0.31.0 contains an uncontrolled resource consumption vulnerability (CWE-400) in its HTTP API request handling logic. Unauthenticated remote attackers can transmit per-request keyword arguments (mm_processor_kwargs and media_io_kwargs) that are passed directly into internal multimodal preprocessing pipelines, causing CPU/GPU memory exhaustion and crashing the engine process.

TL;DR

Unauthenticated remote attackers can crash vLLM multimodal inference instances by supplying extreme keyword parameters in API requests, leading to Out-Of-Memory termination and Denial of Service.

⚠️ Exploit Status: POC

Technical Details

• CWE ID: CWE-400
• Attack Vector: Network (Unauthenticated)
• CVSS v3.1: 7.5 (CVSS:3.1/AV:N/AC:L/PR:N/UI:N/S:U/C:N/I:N/A:H)
• EPSS Score: 0.00343 (26.67th percentile)
• Impact: Denial of Service (Process OOM Crash)
• Exploit Status: Proof-of-Concept / Regression Tests Available
• CISA KEV Status: Not Listed

Affected Systems

• vLLM library versions >= 0.7.3 and < 0.31.0
• Deployments serving MLLMs (Multimodal Large Language Models) via vLLM REST API
• vllm: >= 0.7.3, < 0.31.0 (Fixed in: 0.31.0)

Code Analysis

Commit: 31f2e70

Disable request mm_processor_kwargs and media_io_kwargs by default

Exploit Details

• GitHub Fix Commit: Regression tests included in fix commit demonstrating request validation behavior and OOM prevention.

Mitigation Strategies

• Upgrade vLLM to version 0.31.0 or higher.
• Do not pass the --trust-request-mm-kwargs startup flag in public or multi-tenant environments.
• Filter request parameters at reverse proxies or API gateways to strip mm_processor_kwargs and media_io_kwargs.
• Enforce container memory limits (cgroups) and implement automated process supervision for rapid recovery.

Remediation Steps:

1. Execute pip install --upgrade 'vllm>=0.31.0' in the deployment environment.
2. Inspect server startup flags to ensure --trust-request-mm-kwargs is absent.
3. Restart the vLLM service and confirm HTTP 400 responses when submitting requests containing multimodal kwarg overrides.

References

• GitHub Security Advisory GHSA-gx7p-2j49-hfq4
• Fix Commit 31f2e70cd3e403aaa471fb7fe3d33a7207aaf3cc
• vLLM Pull Request #58830
• vLLM v0.31.0 Release Notes
• NVD Vulnerability Detail - CVE-2026-114055
• CVE.org Vulnerability Record
• VulnCheck Security Advisory

Read the full report for CVE-2026-114055 on our website for more details including interactive diagrams and full exploit analysis.