CVE-2026-105754: vLLM Package Suffers from Scale-Out Transport Vulnerability Leading to Denial of Service
Summary
CVE-2026-105754: vLLM: Scale-out disaggregated multimodal transport trusts caller-supplied features. Affects vllm versions < 0.30.0. Fixed in 0.30.0. How to fix ↓
Check if this vulnerability affects your application
Upload a manifest or SBOM and we'll show whether you're affected, plus every other risky package in the same scan.
No signup required · Results in about 60 seconds
Background
vLLM is an open-source library designed for efficient and scalable multimodal machine learning tasks. It allows users to handle various types of data inputs, such as text and images, and provides a framework for deploying models that can process these inputs in a disaggregated manner. However, a critical vulnerability has been identified in versions up to 0.25.1, where the system improperly trusts caller-supplied features, leading to potential denial of service and data integrity issues.
Vulnerability Detail
The vulnerability arises from vLLM's scale-out transport mechanism, which splits multimodal requests into separate render and generate steps. The generate step accepts a caller-supplied 'features' object without proper validation against the active model's renderer contract. This oversight allows authenticated users to forge requests that can crash the shared EngineCore process, leading to denial of service. Additionally, it can result in silent encoder-cache poisoning and loss of transport integrity, as the system does not verify the integrity of the data being processed.
How To Fix
Latest published version: 0.32.0 as of 22 Oct 2026
pip install --upgrade "vllm==0.32.0"Upgrading to it may include breaking changes; the version below is the minimum fixed release for this vulnerability.
If you need the minimum change
0.30.0
This is the first release that contains the fix.
pip install --upgrade "vllm==0.30.0"Tip — Not sure which version your app runs? Vulert can match this fix to it. Scan free →
Workaround
As a temporary workaround, users can implement additional validation checks in their application code to ensure that the features being passed to the generate step are consistent with the expected model inputs. For example, before processing the features, check the dimensions and types of the input tensors to prevent malformed data from being processed.
Conclusion
The vLLM package has a significant vulnerability that can lead to denial of service and data integrity issues due to improper validation of caller-supplied features. Users are strongly advised to update to the latest version to mitigate these risks and ensure the integrity of their applications.
See every vulnerability recorded for vllm →
FAQ
The vulnerability allows authenticated users to submit forged requests that can crash the EngineCore process or lead to data integrity issues.
You can use Vulert's playground to check if your application is affected by this vulnerability without any signup required.
Implement additional validation checks in your application to ensure that the features passed to the generate step are valid and consistent.
The CVSS score of this vulnerability is 6.5, indicating a moderate severity level.
You can visit the official vLLM GitHub repository for documentation and further details about its functionalities.
References
Related vulnerabilities
Recent vulnerabilities recorded for vllm.
