The vulnerability is an authentication bypass in the AuthenticationMiddleware of vLLM's OpenAI API server. The middleware incorrectly used URL(scope=scope).path to reconstruct the request's URL path for authentication checks. This method is unsafe because it relies on the Host header from the incoming request, which is not properly sanitized by the underlying ASGI server (uvicorn). An attacker could send a request with a crafted Host header (e.g., Host: localhost/v1/models?) to manipulate the reconstructed path. This would cause the authentication check, which verifies if the path starts with /v1, to fail, allowing the request to bypass the API key verification and access protected endpoints. The patch fixes this by using scope["path"] directly, which provides the correct, unalterable request path. The vulnerable function is AuthenticationMiddleware.__call__, as this is where the flawed URL path reconstruction and authentication check occur.