CVE-2026-48746

Modified Modified - Updated After Analysis

Authentication Bypass in vLLM via ASGI Server

Publication date: 2026-06-22

Last updated on: 2026-07-07

Assigner: GitHub, Inc.

Description

vLLM is an inference and serving engine for large language models (LLMs). From 0.3.0 until 0.22.0, a vulnerability in ASGI web servers and starlette's trust on those web servers enables an authentication bypass of the OpenAI API AuthenticationMiddleware. It allows to use the API without providing the configured VLLM_API_KEY or --api-key. This vulnerability is fixed in 0.22.0.

CVSS Scores

EPSS Scores

Probability:
Percentile:

Meta Information

Published

2026-06-22

Last Modified

2026-07-07

Generated

2026-07-13

AI Q&A

2026-06-23

EPSS Evaluated

2026-07-11

NVD

CVE-2026-48746

Affected Vendors & Products

Vendor	Product	Version / Range
vllm	vllm	From 0.3.0 (inc) to 0.22.0 (exc)

Helpful Resources

Exploitability

CWE

KEV

CWE ID	Description
CWE-501	The product mixes trusted and untrusted data in the same data structure or structured message.
CWE-444	The product acts as an intermediary HTTP agent (such as a proxy or firewall) in the data flow between two entities such as a client and server, but it does not interpret malformed HTTP requests or responses in ways that are consistent with how the messages will be processed by those entities that are at the ultimate destination.

Attack-Flow Graph

Impact Analysis

The vulnerability allows unauthorized users to access the API without authentication, which can lead to unauthorized use of the service. Given the CVSS score of 9.1 with high confidentiality and availability impacts, this could result in data exposure or denial of service, potentially disrupting operations or exposing sensitive information.

Mitigation Strategies

To mitigate this vulnerability, upgrade vLLM to version 0.22.0 or later, where the authentication bypass issue has been fixed.

Executive Summary

This vulnerability affects vLLM, an inference and serving engine for large language models. Specifically, from versions 0.3.0 until 0.22.0, a flaw in ASGI web servers and the way starlette trusts those servers allows an authentication bypass in the OpenAI API AuthenticationMiddleware. This means that an attacker can use the API without providing the required VLLM_API_KEY or --api-key.

Compliance Impact

This vulnerability allows authentication bypass of the OpenAI API AuthenticationMiddleware in vLLM, enabling unauthorized use of the API without the required API key.

Such unauthorized access could lead to exposure or misuse of sensitive data processed by the API, potentially violating data protection requirements under standards like GDPR or HIPAA.

Therefore, this vulnerability may negatively impact compliance with regulations that mandate strict access controls and protection of sensitive information.

Detection Guidance

This vulnerability involves an authentication bypass in the OpenAI API AuthenticationMiddleware of vLLM versions from 0.3.0 until 0.22.0, caused by improper path normalization in ASGI web servers and starlette.

To detect if your system is vulnerable, you can attempt to access the API endpoints without providing the configured VLLM_API_KEY or --api-key and observe if access is granted.

Specifically, you can test for bypass by sending HTTP requests with multiple leading slashes in the path to see if the authentication middleware improperly allows access.

Example command using curl to test the vulnerability:

curl -i -H "Authorization:" http://your-vllm-server//api/endpoint

If the server responds with successful access (e.g., HTTP 200) despite missing or empty Authorization headers, it indicates the vulnerability is present.

Note that the vulnerability is fixed in vLLM version 0.22.0, so upgrading is the recommended mitigation.

Hi! I’m here to help you understand CVE-2026-48746. Ask me anything about the vulnerability, its impact, or mitigation strategies.

0/70