Key points
- vLLM versions before 0.31.0 are vulnerable to uncontrolled resource consumption.
- Attackers can manipulate processor parameters to crash shared servers.
- Version 0.31.0 is available and resolves the CVE-2026-114055 issue.
Users running vLLM versions prior to 0.31.0 face a high-severity security risk that allows remote attackers to disrupt service. The vulnerability, tracked as CVE-2026-114055, stems from how the software handles multimodal processing parameters. According to the National Vulnerability Database, this flaw enables unauthorized control over resource usage, leading to server termination.
In plain English
The vLLM framework allows applications to process images and videos alongside text. In affected versions, the system trusts input values for processing tasks without sufficient limits. An attacker can send requests that specify extremely large image sizes or excessive frame counts. This forces the server to allocate massive amounts of memory, causing it to crash or become unresponsive.
The background
The issue exists in vLLM releases starting from version 0.7.3 up to, but not including, version 0.31.0. The flaw is classified as uncontrolled resource consumption under CWE-400. It specifically impacts multimodal processors such as Isaac, InternVL, and H2OVL. By supplying custom parameters in the `mm_processor_kwargs` field, remote clients can dictate how much memory the processor uses per request.
What changes now
The vLLM project has released version 0.31.0, which addresses this vulnerability. Organizations using earlier versions should upgrade immediately to prevent service disruption. The fix restricts how processor parameters are handled, ensuring that malicious inputs cannot exhaust system resources. This update is critical for any shared server environment running multimodal models.
Who needs to act first
Teams should prioritize updating systems that expose vLLM services to untrusted networks or the public internet. Before applying the update, verify that current workloads are backed up and that the new version is compatible with existing applications. Test the upgrade in a staging environment to ensure stability. If immediate patching is not possible, restrict access to the service using network-level controls until the fix can be deployed.
What to do and how to stay safe: vLLM
- Upgrade vLLM to version 0.31.0 or later on all affected servers.
- Review network access controls to limit who can send requests to multimodal endpoints.
- Monitor system memory usage for unusual spikes that may indicate exploitation attempts.
- Verify that all dependent applications function correctly after the update.
Step-by-step guide: Out-of-Band Patches: Meaning, Risks, and Deployment Reality
General security guidance from the Patch Herald newsroom. It is not confirmed advice from the organisations named in this story.
Frequently asked questions
Which vLLM versions are affected by CVE-2026-114055?
Versions from 0.7.3 up to, but not including, 0.31.0 are vulnerable.
Is there a fix available for this vulnerability?
Yes, upgrading to vLLM version 0.31.0 resolves the issue.
What is the severity of this vulnerability?
The National Vulnerability Database rates it as HIGH with a CVSS score of 8.7.
