Aidan Finn, IT Pro

I see many people implementing classic network security designs in Azure. Maybe there’s DMZ and an internal virtual network. Maybe they split Production, Test, and Dev into three virtual networks. Possibly, they do a common government implementation – what Norway calls “Secure Zone”. I’m going to explain to you why these network designs offer very little security.

I have written this post as a contribution to Azure Spring Clean 2025. Please head over and check out the other content.

Essential Reading

This post is part of a series that I’ve been writing over several weeks. If you have not read my previous posts then I recommend that you do. I can tell that many people assume certain things about Azure network based on designs that I have witnessed. You must understand the “how does it really work” stuff before you go any further.

• Azure Virtual Networks Do Not Exist
• How Do Network Security Groups Work?
• Beware Of The Default Rules In Network Security Groups
• How Many Subnets Do I Need In An Azure Virtual Network?
• Routing Is The Security Cabling of Azure
• How Does Azure Routing Work
• How Many Azure Route Tables Should I Have?
• Micro-Segmentation Security In Azure Networks

A Typical Azure Network Design

Most of the designs that I have encountered in Azure, in my day job and as a community person who “gets around”, are very much driven by on-premises network designs. Two exceptions are:

• What I see produced by my colleagues at work.
• Those using Enterprise Scale from the Microsoft Cloud Adoption Framework – not that I recommend implementing this, but that’s a whole other conversation!

What I mostly observe is what I like to call “big VNets”. The customer will call it lots of different things but it essentially boils down to a hub-and-spoke design that features a few large virtual networks that are logically named:

• Dev, Test, and Production
• DMZ and private
• Internal and Secure

Workload: A collection of resources that provide a service. For example, an App Service, some Functions, a Redis cache, and a database might make a retail system. The collection of resources is a workload, united in their task to provide a service for the organisation.

You get the idea. There are a few spoke virtual networks that are each peered to a hub.

The hub is a transit network, enabling connectivity between each of the big VNets – or “isolating them completely” – except for where they don’t (quite real, thanks to business-required integrations or making the transition from testing to production easier for developers). The hub provides routing to Azure/The Internet and to remote locations via site-to-site networking.

If we drill down into the logical design we can see the many subnets in each spoke virtual network. Those subnets are logically divided in some way. Some might do it based on security zones – they don’t understand NSGs. Some might have one subnet per workload – they don’t know that subnets do not exist. Each subnet has an NSG and a Route Table. The NSG “micro-segments” the subnet. The Route Table forces traffic from the subnet to the firewall – the logic here can vary.

Routing & Subnet Design

Remember three things for me:

• Virtual networks and subnets do not exist – packets go directly from sender to receiver in the software-defined network.
• Routing is our cabling when designing network security.
• The year is 2025, not 2003 (before Windows XP Service Pack 2 introduced Windows Firewall to the world).

There might be two intents for routing in the legacy design:

• Each virtual network will be isolated from the others via the hub firewall.
• Each subnet will be isolated from the others via the hub firewall.

Big VNet Network Isolation

Do you remember 2003? Kid Rock and Sheryl Crow still sang to each other. Avril Lavigne was relevant (Canada, you’re not getting out of this!). The Black Eyed Peas wanted to know where the love was because malware was running wild on vulnerable Windows networks.

I remember a Microsoft security expert wandering around a TechEd Europe hall, shouting at us that network security was something that had to be done throughout the network. The edge firewall was like the shell of an egg – once you got inside (and it didn’t matter how) then you had all that gooey goodness without any barriers.

A year later, Microsoft released Windows XP/Windows Server 2003 Service Pack 2 to general availability. This was such a rewrite that many considered it a new OS, not a Service Pack – what the kids today call a feature update, a cumulative update, or an annual release. One of the new features was Windows Firewall, which was on by default and blocked stuff from getting into our machines unless we wanted that stuff. And what did every Windows admin do? They used Group Policy to turn Windows Firewall off in the network. So malware continued, became more professional, and became ransomware.

Folks, it’s been 21 years. It’s time to harden those networks – let the firewall do what it can do and micro-segment those networks. Microsoft tells you to do it. The US NSA tells you to do it. The Canadian Centre for Cyber Security tells you to do it. The UK NCSC tells you to do it. Maybe, just maybe, they know more about this stuff than those of you who like gooey network insides?

Big VNet Subnet Isolation

The goal here is to force any traffic that is leaving a subnet to use the hub firewall as the next hop. In my below example, if traffic wants to get from Subnet 1 to Subnet 2, it must first pass through the firewall in the hub. A Route Table is created with a collection of User-Defined Routes (UDR) such as shown below.

Each UDR uses Longest Prefix Match to force traffic to other subnets to route via the firewall. You don’t see it in the diagram, but there would also be a route to 0.0.0.0/0 via the firewall, including any prefix outside of this virtual network, except the hub (Longest Prefix Match selecting the System route created by peering with the hub).

Along comes the business and they demand another workload or whatever. A new subnet is required. So you add that subnet. It’s been a rough Friday and the demand came right before you went home. You weren’t thinking straight and .. hmm … maybe you forgot to update the routing.

Oh it’s only one Route Table for Subnet 4, right? Em, no; you do need to add a route table to Subnet 4 with prefixes to subnets 1-3 and 0.0.0.0/0. But that only affects traffic leaving Subnet 4.

What you forget is that routing works in two ways. Subnets 1-3 require a UDR each for Subnet 4, otherwise traffic from Subnets 1-3 will route directly to Subnet 4 and the deeper inspection of the firewall won’t see the traffic. Worse, you probably broke TCP communications because you set up an asynchronous route and the stateful hub firewall will block responses from Subnet 4 to Subnets 1-3.

Imagine this Production VNet with 20, 30, or 100 subnets. This routing update is going to be like like manual patching – which never happens.

One of the biggest lessons I can share in secure network design is KISS: keep it simple, stupid. Routing should be simple, and routing should be predictable when there is expansion/change, because routing is your cabling for enforcing or bypassing network security.

Network Security Group Design

As a consultant, I often have a kickoff meeting with a customer where they stress how important security is. I agree – it’s critical. And then I get to see their network or their plans. At this point, I shouldn’t be surprised but I always am. Some “expert” who passed an Azure certifcation exam or three implements a big VNet design. And the NSGs – wow!

What you’ll observe is:

• They implement subnets as security zones, when the only security zoning in Azure is the NSG. NSG rules, processed on the NIC, are how we allow/deny incoming or outgoing traffic at the most basic level. In the end, there are too many subnets in an already crowded big VNet.
• The NSG either uses lots of * (any) in the sources and destinations leading to all sorts of traffic being allowed from many locations.
• They think that they are blocking all incoming traffic by default but don’t understand what the default rule 65000 does – it lets every routable network (Azure & remote) in.
• They open up all traffic inside the subnet – who cares if some malware gets in via devops or a consultant who uploads it via a copy/paste in RDP?

And they’ll continue to stress the importance of security.

Shared Resources In The Hub

This one makes me want to scream. And to be fair, Microsoft play a role in encouraging this madness – shame on you, Microsoft!

The only things that should be in your hub are:

• Virtual Network Gateways
• Third-party routers and Azure Route Server
• The firewall
• Maybe a shared Azure Bastion with appropriate minmised RBAC rights

That’s it! Nothing else!

Don’t put DNS servers here. Don’t put a “really important database” in the hub. Don’t put domain controllers in the hub. Repeat after me:

I will not place shared resources in the hub

Everything is a shared resource. Just about every workload shares with other workloads. Should all shared resources go in the hub? What goes in the spokes now?

“Why?” you may ask. Remember:

• By default, everything goes straight from source to destination
• Routing is our way to force traffic through a firewall
• When you peer two VNets, a new System route enables direct connectivity between NICs in the two VNets.

People assume that a 0.0.0.0/0 route includes everything, but Longest Prefix Match overrides that route when other routes exist. So, if you place a critical database in the hub, spokes will have direct connectivity to that database without going through the firewall and any advanced inspection/filtering services that it can offer – and vice versa. In other words:

• You opened up every port on the critical resource to every resource in every spoke.
• You created an open bridge between every spoke.

And the fact is that putting something in the hub doesn’t make it “more shared” (how is it less shared than something in a spoke?) or faster (software-defined networking treats two NICs in peered VNets as if they were in the same VNet).

Those clinging to putting things in the hub will then want more routes and more complexity. What happens when the organisation goes international and adds hub & spoke deployments in other regions? What should be a simple “1 peering & 1 route” solution between two hubs will expand into routes for each hub subnet containing compute.

Everything is shared – that’s modern computing. Place your workloads into spokes, whether they are file shares, databases, domain controllers, or DNS servers/Private Resolvers. They will work perfectly well and your network will function, be more secure, simpler to manage/understand, and the security model will be more predictable.

Wrapping Up

This is a long post. There is a good chance that I just spat in the face of your cute lil’ baby Azure network. I will be showing you alterantives in future posts, building up the solution a little at a time. Until then, KISS … keep it simple, stupid!

Here comes yet another “How does it work” post on Azure networking. I have observed many folks who assume that routing in Azure works one way, but are shocked to learn that there are more layers than they anticipated. In this post, I will explain how routing really works in Azure networking.

The Misconception

I will start by revisiting a Microsoft diagram that I previously used for a discussion on the importance of routing in network security.

The challenge with the above architecture is to make traffic flow through the firewall. Most people will answer that User-Defined Routes (UDRs) via Route Tables are required. Yes, that is true. But they fail to understand that two (I would argue three) other sources of routes are also present in this diagram. The lack of that additional knowledge may impact this simple scenario. And I know for certain that if this scenario were the typical mid-large organisation, then the lack of knowledge would become:

• An operational issue
• A security issue
• A troubleshooting issue
• A connectivity issue

The NIC Is The Router

One of my first posts in this series was “Azure Virtual Networks Do Not Exist“. In that post, I explained that all traffic routes directly from the source NIC to the destination NIC. There is no subnet, no default gateway, and no virtual network. Instead, a virtual network is a mapping of a mesh connectivity between all NICs in that virtual network. When you peer virtual networks, the mapping expands to mesh all NICs in the peered virtual networks.

Where does routing happen if there is no default gateway or subnet? The answer (just like “where are NSG rules processed?” is the NIC is the router.

Remember that everything is a virtual machine, including “serverless computing”, somewhere in the platform.

If packets travel directly from source to destination, then there is no router appliance between the source and the destination. That means that the source must be its own router.

Some Basic Routing Theory

A route is an instruction: if you want to get to address A then go to place X. X might be the destination, or it might be the first hop to get to the destination.

For example, I might have a remote network of 192.168.0.0/16. I have an Azure App Service that wants to use a site-to-site connection to reach out to a server with an address of 192.168.1.10. A route might say:

• Prefix: 192.168.0.0/16
• Next Hop Type: Virtual Network Gateway (VPN or ExpressRoute)

The NIC of the App Service will learn that route (see BGP later). Packets from the App Service will go directly to the NIC(s) of the Virtual Network Gateway and then route over VPN/ExpressRoute to 192.168.1.10.

Maybe I will manipulate that route a little to force egress traffic through a firewall. My firewall will have an internal IP address of 10.0.1.4. I can introduce a route (see User-Defined Routes later) of:

• Prefix: 192.168.0.0/16
• Next Hop Type: Virtual Appliance
• Next Hop IP Address: 10.0.1.4

Now packets to 192.168.1.10 will go to my firewall. It’s important now that the firewall has a route to 192.168.0.0/16 – normally it would by default in a hub & spoke design.

The second piece of knowledge to have is that there must be a route for the response. There is no implied return route. Either a human or the network must implement that return route. And it’s really important that the return route is the same as the egress route; stateful firewalls will block TCP responses when they have not permitted the requests – this is one of those “you’ll learn it the hard way” things when dealing with site-to-site connections and firewalls.

The Laws Of Azure Routing

I will revisit this at the end, but here’s what you need to know when you are designing/troubleshooting routing in Azure:

1. Route source priority
2. Longest prefix match

Law 1: Route Source Priority

You might know that User-Defined Routes (UDRs) exist. But there are two (or three) other sources of routes and they each have a priority.

System Routes

The first source of routes that is always there is System (or Default) routes. System routes are created when you create or configure a virtual network. For example, every subnet in a brand-new virtual network has many system routes out of the box. The major routes we are concerned with are:

• Route(s) to the address prefix(es) of the virtual network to route directly (VirtualNetwork) to the destination NICs.
• A route to send all other traffic to the Internet (including Azure).

Yes, I am leaving out a bunch of other system routes that are implemented to protect Microsoft 365 from hacking but I want to keep this simple.

Another important System route is what is created when you peer two virtual networks. A route is created in each of the peered virtual networks to state that the next hop to the new neighbour is via peering. This is a human-friendly message; what it means is that the NICs in the connected peer are now part of the local virtual network’s mesh – packets from local NICs will route directly to NICs in the peered virtual network.

BGP Routes

Border Gateway Protocol (BGP) is a mechanism where one routing appliance shares its knowledge of routes with neighbours. For example, a router in Dublin might say “If you want to get to any NICs in Dublin then come to me”. A router in Paris might hear that message and relay it by saying “I know how to get to Dublin so if you want to get to Dublin, come to me”. A router in Munich might pick up that relay from Paris and advertise locally that it knows how to get to Dublin. A PC in Munich wants to send a packet to a NIC in Dublin. The Munich network says that the route to Dublin is via the router in Munich, so the flow of packets will be:

Munich PC > Munich router > Paris router > Dublin Router > Dublin IP NIC

Azure implements BGP in two scenarios:

• Site-to-site networking
• Azure Route Server

You must configure BGP when using ExpressRoute for remote site connections. You optionally configure BGP when configuring a BGP tunnel. What most people don’t realise is that you will still have BGP routes with a BGP-less VPN tunnel thanks to the Local Network Gateway which generates BGP routes for the remote site prefixes. In the case of site-to-site networking, BGP routes are propagated from the GatewaySubnet and propagate to all other subnets in the virtual network and (by default) to all peered virtual networks/subnets.

The other scenario is Azure Route Server (ARS), which also includes Virtual WAN, where the router is Azure Route Server – Azure Route Server originated in Virtual WAN. ARS can peer with other appliances, such as a router Network Virtual Appliance (NVA), and share routes with it:

• Routes of remote connected networks are learned from the NVA and propagated to the Azure hub/spokes. The hub/spokes now know that the route to the remote networks is to use the router as the next hop (not your firewall!).
• The prefixes of the hub/spokes are shared with the NVA to enable remote networks to know how to get to them.

User-Defined Routes (UDRs)

This is the one kind of route that we can directly manage as Azure architects/administrators/operators. A resource called a Route Table is created. The Route Table is associated with a subnet and applies its settings to all NICs in the subnet. There are two important things we can use the Route Table for:

• Disable BGP Propagation: We can disable inward BGP route propagation to the associated subnet. This means that we can prevent routes to remote sites from bypassing our firewall by using the Virtual Network Gateway/NVA as the next hop.
• User-Defined Routes: We can implement routes that force traffic in ways that we want.

UDRs have several possible next hops for packets:

• Virtual Appliance: A router or firewall – you additionally specify the IP address of the virtual appliance NIC to use.
• Internet: Including the Internet and Azure
• Virtual Network Gateway: An Azure site-to-site connection in the virtual network or shared with the virtual network via peering.
• Virtual Network: Send packets to the same virtual network.
• None: The packets are dropped at the source NIC and are never transmitted – a useful security feature.

Hidden Programmed Routes

You won’t find this one in any official documentation on routing but it does exist and you’ll learn about them either by accident or by educated observation of behaviour.

Microsoft will sometimes introduce a system route to fix an issue where if you do X, they will program a route to be generated. Unfortunately, this (probably a) type of System route cannot be visibly observed in any way because no diagnostics tools exist for that subnet.

One example of this is Private Endpoint. When you create a subnet, network policies for Private Endpoint are disabled by default. This causes a chain of things to happen:

• UDRs are ignored by Private Endpoints in the subnet
• Each Private Endpoint in the subnet will create its own /32 (the IP address of the Private Endpoint is the destination prefix) System Route in the virtual network and directly peered virtual networks. This means that a /32 route for the Private Endpoint is added to the GatewaySubnet of the hub/spoke depending on your design.

That GatewaySubnet System route has broken the spirit of many Azure admins over the years. You can’t see it and, from our perspective, it shouldn’t exist. The result was that traffic from on-premises to Private Endpoints went directly to the Private Endpoint, even if we set up a UDR to force traffic to the spoke virtual network to go via the firewall. This is because of the second law of routing: Longest Prefix Match.

Route Deactivation

We have established that there are three^* sources of routes. What happens if two or three of them create routes to the same prefix? That can happen; in fact, you will probably make it happen if you want to force traffic through a firewall.

Let’s imagine a scenario where there are 3 routes to 192.168.0.0/16 from:

• System
• BGP
• UDR

What happens? The fabric handles this automatically and applies a prioritisation rule to deactive the routes from lesser sources. The priority is as follows:

1. UDR: Routes that you explicity create in Azure will deactive routes from BGP & System to the same prefix. UDR beats BGP & System.
2. BGP: Routes that are created by admins/networks in other locations will deactivate routes from System to the same prefix. BGP beats System.
3. System: System routes are Azure generated and get beat by BGP and UDR routes to the same prefix.

Let’s consider a simple/common example. We have a virtual network with a subnet. If you want to see this in action, add a VM to the subnet, power it up, open the Azure NIC resource, and go to Effective Routes (wait 30 seconds). Withotu doing anything to the subnet/virtual network a System Route will be created for all NICs in the subnet:

• Prefix: 0.0.0.0/0
• Next Hop Type: Internet

What that means is that any traffic that doesn’t have a route will be sent to Internet.

Let’s say that I want to force that traffic through a firewall appliance with an IP address of 10.0.1.4. I can associate a new Route Table to the subnet and add a UDR to the subnet:

• Prefix: 0.0.0.0/0
• Next Hop Type: Virtual Appliance
• Next Hop IP Address: 10.0.1.4

Two routes to 0.0.0.0/0 are present. Which one will be used? That decision is already made. The System route to 0.0.0.0/0 is automatically deactivated by the fabric as soon as a higher (BGP or UDR) route is added to the subnet. The only active route to 0.0.0.0/0 in that subnet is my UDR via the firewall.

Law 2: Longest Prefix Match

There is another scenario where there may be multiple route options. A packet might be destined to an IP address and multiple active routes might be applicable. In this case, Azure applies “Longest Prefix Match” – you can think of it as the best matching route. This one is best explained with an example.

Let’s say a packet is going 10.10.10.4. However, the source NIC has 3 possible routes that could apply:

• System: 0.0.0.0/ via Internet
• BGP: 10.10.10.0/24 via Virtual Network Gateway
• UDR: 10.0.0.0/8 via a firewall

All of the routes are active because the prefixes are different. Which one is chosen? Tip: Route priority (UDR/BGP/System) is irrelevant now.

I don’t know the internal mechanics of this but I suspect that an AND operation is done using the destination address and the route prefix. Remember that each octet in a 32 bit IP address is 8 bits:

Here is the calculation for the System route, which sums to 0 bits:

Route Prefix	0	0	0	0
Destination	10	10	10	4
AND Bits	0	0	0	0

Here is the calculation for the BGP route, which sums to 24 bits:

Route Prefix	10	10	10	0
Destination	10	10	10	4
AND Bits	8	8	8	0

Here is the calculation for the UDR route, which sums to 8 bits:

Route Prefix	10	0	0	0
Destination	10	10	10	4
AND Bits	8	0	0	0

Which route is the best match? The BGP route is because it has the longest prefix match to the destination IP address.

Review: The Laws of Azure Routing

Now you’ve learned how Azure routes are generated, how they are prioritised, and how they are chosen when a packet is sent. Let’s summarise the laws of Azure routing:

1. Route Source Priority: When there are routes to the same prefix, BGP beats Sytem, and UDR beats BGP & System.
2. Longest Prefix Match: When multiple routes can be used to send a packet to a destination, the route with the longest bit match will be selected.
3. It’s Always DNS: Ask any Windows admin – when routing isn’t the cause of issues, then it’s DNS 🙂

You’re designing a new virtual network in Azure. You’re going to have three different security zones in your application. How many subnets do you need? I will help you understand why many of you gave the incorrect answer.

Back To Basics

In a previous post, I explained that virtual networks do not exist. Therefore, subnets do not exist. That’s why you cannot ping a default gateway. Packets do not leave a source NIC and route via default gateway to enter another subnet. Packets go from the source NIC, disappear in the physical network of Azure, and reappear at the destination NIC, whether it is on the same host, in the same data centre, in a neighbouring data centre, or on the other side of the world. Say it after me:

Subnets do not exist.

If packets go straight from source to destination, what is the logic of creating subnets to isolate resources?

Why Did We Segment Networks Using Subnets?

In the on-premises world, there are many reasons to segment a network. A common reason was to control the size of broadcast/multicast domains. That’s not an issue in Azure because virtual networks do not support broadcasts/multicasts.

From a security perspective, we segmented networks because we needed to isolate a firewall. The firewall is a central resource. A network runs from a top-of-rack switch to an ethernet interface in the firewall. That subnet uses the firewall to route to other subnets, possibly using the same cable (VLANs) or using different cables to other top-of-rack switches.

Earlier I asked you to imagine a workload with three security zones. Let’s call them:

• Web
• Application
• Database

That’s not too crazy. My security model requires me to ensure:

• Internet users can only reach the web servers on HTTPS
• The Application server can only be talked to by the web servers.
• The database servers can only be talked to by the application servers.

How would I create that? I’d set up three VLANs or subnets. Each VLAN would use a default gateway which is either the firewall or uses the firewall as a next hop to reach other VLANs. The firewall would then enforce my security intent, ensuring that only desired traffic could enter a VLAN to reach the required machines.

This design works perfectly well in on-premises cable-oriented networks because the networks (physical or virtual) are connected via cable(s) running to the firewall.

Bringing Cable-Oriented Designs To Azure

There is no finger-pointing here – I still have nightmares about an early Azure design I did where I created a VNet diagram with somewhere between 10-20 subnets. We all learn, and I’m hoping you learn from my mistakes.

Using the same requirements as before for our workload, we can produce the below design … based on cable-oriented patterns.

We create a single virtual network broken into 3 subnets. Each subnet has VMs for each role in the application. We then isolate each of the machines using NSGs.

That seems perfect, right? It is secure. Traffic will get from A to B. If we implement the rules correctly, then only the correct traffic will flow. But this design does display a lack of understanding.

Remember: packets go directly from source to destination. There is no default gateway. If an NSG that is processing rules on an Application Server NIC is allowing or denying traffic, then what is the point of the subnet? The subnet is not doing the segmentation; the NSG is doing the segmentation.

How Can We Segment Networks In Azure?

The most basic segmentation method in an Azure virtual network is the Network Security Group (NSG). While the previous Azure diagram is not technically wrong, the below diagram displays a better understanding of the underlying technology:

In this design, we are accepting that neither the virtual network nor the subnet exist. We are using rules in the NSG to isolate each tier of the application:

Look at the below NSG to see how this isolation can be done with a very simple example:

The NSG denies all traffic by default (rule 4000). Then the only traffic permitted is what we modeled previously using subnets. The rules are processed on the NICs, so the only way traffic enters a VM is if it is compliant with the above NSG.

Yes, I could use groups of IP addresses, or better, Application Security Groups that make the rules more readable and allow aggregation/abstraction of NICs & IP addresses.

So Why Do We Create Subnets In Azure

The primary reason is quite boring: technical requirements. Let me adjust my design a little. The database is going to be implemented using SQL Managed Instance instead of a VM. In the original VM-only design, there were no impediments to using a single subnet. SQL Managed Instance changes the technical requirements because it must be connected to a dedicated subnet.

That’s a simple example. A different example might be that I must use different address prefixes – see an older post by me on using a Linux VM as a NAT gateway where the VM has an internal NIC on a regularly addressed subnet and a second NIC in a subnet that is addressed based on NAT requirements.

Another example might be that you need to create custom routes for different NICs to the same prefix. For example, some NICs will go via your firewall to 0.0.0.0/0. Other NICs might go to “None” (a blackhole that drops packets) for traffic going to 0.0.0.0/0. The only way to implement that is to have one subnet for each Route Table. I’m not going to dive into routing here – let’s save that for another day.

Taking This Bigger

I am eventually going to explain enough things so I can show you why the classic Azure “big VNet” likely called production, test, or dev, is both an operational and security nightmare. But the above content, along with my other recent posts, are just part of the puzzle. Watch out for more content coming soon.