/

CVE-2026-48746

Severity:

CRITICAL RISK

(9.1)

Auth Bypass (CWE-444) in vLLM OpenAI API Middleware

CVE-2026-48746

Modified

Vulnerability type

Category:

Authentication Bypass / Improper Access Control

CWE:

CWE-444, CWE-501

Impact:

Unauthorized API Access and Abuse of LLM Inference Services

Severity:

CRITICAL RISK

(9.1)

CVSS version:

3.1

Catch this CVE while code is being written

Detect vulnerable code directly in pull requests, so CVEs are fixed before merge.

Summary

An authentication bypass in vLLM's OpenAI-compatible API allows requests to be accepted without the configured VLLM_API_KEY or --api-key being presented.

Under certain ASGI web server deployments where Starlette overly trusts upstream headers or connection metadata, attackers can call the API as if properly authenticated.

Affected context

The vulnerability occurs in vLLM deployments exposing the OpenAI-compatible HTTP API through ASGI web servers where Starlette's assumptions about trusted request metadata are not aligned with the actual network or proxy topology.

It typically manifests when upstream servers, reverse proxies, or load balancers can influence headers or connection attributes that the AuthenticationMiddleware relies on to decide whether a request is authenticated.

In such environments, an attacker sending crafted requests to the public-facing endpoint may be treated as authenticated even without providing the configured API key.

Exploit preconditions

  • Attack surface: Public or internal vLLM OpenAI-compatible HTTP endpoints reachable over the network.
  • Required access: Network access to the ASGI-served vLLM API; no valid API key is required if the deployment is misconfigured or implicitly trusts upstream metadata.
  • Trigger point: HTTP requests that omit or manipulate authentication headers or parameters which the AuthenticationMiddleware expects (e.g., missing or invalid API key).
  • Why exploitation is possible: The middleware and framework logic rely on trust assumptions about the ASGI server and request metadata that can be influenced or spoofed, leading to a failure to enforce the intended API key check.

Why this happens

This issue stems from a mismatch between trust boundaries and how authentication decisions are made in the middleware stack.

The OpenAI API AuthenticationMiddleware assumes that certain request attributes or headers provided by the ASGI server are authoritative indicators of client authenticity.

When those assumptions are incorrect—such as when upstream servers or clients can control values that are implicitly trusted—the logic can skip or incorrectly interpret API key checks.

At an abstract level, this aligns with CWE-444: Inconsistent Interpretation of HTTP Requests, where different components in the request handling chain (proxies, ASGI server, application middleware) interpret or trust HTTP semantics differently, creating gaps attackers can exploit to bypass authentication.

Potential impact

  • Attackers may invoke LLM inference endpoints without a valid API key.
  • Unmetered or abusive consumption of compute resources and GPU capacity.
  • Potential exposure of model behaviors, system prompts, or tuned configurations through unauthorized querying.
  • Financial impact due to resource exhaustion or unexpected usage in metered environments.
  • Denial of service for legitimate users if infrastructure is saturated by unauthorized traffic.

General mitigation guidance

Mitigation focuses on tightening authentication boundaries and aligning trust assumptions across the entire HTTP handling path.

Useful defensive patterns include:

  • Designing authentication logic so that only explicit, application-controlled credentials (e.g., API key headers or tokens) are used to grant access, independent of upstream connection metadata.
  • Avoiding implicit trust in headers or attributes that can be influenced by clients or intermediate proxies, unless they are clearly within a trusted network boundary and protected by strict controls.
  • Enforcing a single, clear source of truth for authentication decisions in middleware, and failing closed when expected credentials are missing or malformed.
  • Segmenting vLLM inference endpoints behind authenticated gateways or API management layers that consistently enforce key or token validation.
  • Applying least-privilege principles to accessible endpoints and separating internal management APIs from public inference APIs.

How teams detect issues like this

Detection can focus on both code-level patterns and runtime behavior:

  • Static analysis or code review to identify middleware that conditionally skips authentication based on headers, connection info, or assumptions about the ASGI server.
  • Policy checks to ensure all external-facing endpoints pass through a common, consistently enforced authentication layer that requires explicit credentials.
  • Log analysis and anomaly detection to flag requests that are treated as authenticated despite missing or invalid API keys, particularly from untrusted networks.
  • Security testing and fuzzing of HTTP header combinations, proxy configurations, and authentication flows to uncover inconsistencies between expected and actual enforcement.
  • Use of automated security review tools that scan for routes or middleware paths where authentication can be bypassed under specific server or proxy configurations.

Sources

SourceLink
github.comgithub.com/vllm-project/vllm/pull/43426
github.comgithub.com/vllm-project/vllm/security/advisories/GHSA-94f4-hr76-p5j6
x41-dsec.dex41-dsec.de/lab/advisories/x41-2026-002-starlette
access.redhat.comaccess.redhat.com/errata/RHSA-2026:30088
access.redhat.comaccess.redhat.com/errata/RHSA-2026:30089
access.redhat.comaccess.redhat.com/errata/RHSA-2026:36005
access.redhat.comaccess.redhat.com/errata/RHSA-2026:36006
access.redhat.comaccess.redhat.com/errata/RHSA-2026:42132
access.redhat.comaccess.redhat.com/errata/RHSA-2026:42142
access.redhat.comaccess.redhat.com/errata/RHSA-2026:42644
access.redhat.comaccess.redhat.com/errata/RHSA-2026:43038
access.redhat.comaccess.redhat.com/errata/RHSA-2026:61627
access.redhat.comaccess.redhat.com/errata/RHSA-2026:61629
access.redhat.comaccess.redhat.com/security/cve/CVE-2026-48746
bugzilla.redhat.combugzilla.redhat.com/show_bug.cgi?id=2491581
security.access.redhat.comsecurity.access.redhat.com/data/csaf/v2/vex/2026/cve-2026-48746.json

Frequently asked

questions

Yes. If the vLLM OpenAI-compatible API is reachable over the network and deployed with affected trust assumptions, an attacker can exploit this issue remotely by sending crafted HTTP requests without a valid API key.

The attacker typically only needs network access to the exposed vLLM endpoint. They do not require a valid API key; exploitation relies on how the ASGI server and Starlette interpret and trust HTTP request metadata.

Centralize authentication in a strict middleware layer that always requires explicit, application-validated credentials, and make it independent of proxy- or server-supplied indicators that could be influenced by clients.

Ensure the logic fails closed when API keys or tokens are missing or invalid, and avoid configuration modes where upstream headers can change authentication decisions.

When designing APIs, define clear trust boundaries and document which components are responsible for authentication.

Use defense-in-depth: combine strong, explicit credential checks in the application with secure API gateways or reverse proxies, test deployments under realistic proxy and header conditions, and include authentication-flow tests in CI to catch bypasses caused by configuration or framework changes.

Find out what's already
exploitable in your codebase.