Because an ERP report cannot pull data from the historian, someone opens a “temporary” rule from IT to OT; months later, that rule is still there. The OT DMZ is the intermediate layer between the enterprise network (Levels 4–5) and the production network (Levels 0–3), known as Level 3.5 in the Purdue model, and it exists precisely to close off shortcuts like this. Its job is not to connect the two networks but to terminate every data exchange between them at an intermediary server. Below, for IT/OT teams building a DMZ from scratch or overhauling an existing one, we describe in the order used on site which servers go into the DMZ, which side initiates each flow, and how the rule base is written and audited.
The purpose of the OT DMZ and the ban on direct sessions
The rule is one sentence: no direct TCP/UDP session is established between a system on the IT network and a system on the OT network. Every flow terminates at a server in the DMZ; the other side receives the data from that server over a separate session. If the ERP wants production data, it connects not to the OT historian but to the historian replica in the DMZ. If an engineer needs to access a PLC project from home, they first go to the jump host in the DMZ and from there, over a separate session, to the engineering workstation in OT. There are no exceptions.
In practice, this means there is not a single allow line of the form “source IT, destination OT” in the firewall rule base. This is the first place we look when auditing an existing DMZ; a common finding is exactly such a line, opened during commissioning and forgotten. For the general framework of the Purdue levels, see the article IT/OT network separation with the Purdue model; here we cover only the internal design of Level 3.5.
The second rule is flow direction. Flows initiated from inside OT towards the DMZ (push) are preferred over those initiated from the DMZ towards OT (pull). This keeps the number of inbound connection rules on the OT-side firewall to a minimum, and even if a server in the DMZ is compromised, its ability to open new sessions into OT stays narrow. Keep inbound rules few.
Single-firewall and dual-firewall models
A DMZ is built with one of two topologies. In the single-firewall model, one interface of a three-legged device faces IT, one the DMZ and one OT. In the dual-firewall model, separate devices sit at the IT–DMZ and DMZ–OT boundaries, often managed by different teams.
| Criterion | Single firewall (three-legged) | Dual firewall |
|---|---|---|
| Cost and operation | Low, single rule base | Higher, two rule bases |
| Single point of failure | One configuration error can open both boundaries | An error on one device affects a single boundary |
| Separation of management | IT and OT manage the same device | IT can manage the outer firewall, OT the inner one |
| Where it fits | Small to medium plants, limited number of flows | Critical processes, many flows, a separate OT team |
In the single-firewall model, most of the risk lies in rule writing. A single line containing “any” between the IT and OT interfaces renders the DMZ useless. If you chose this model, put an explicit “deny all” rule for the IT→OT and OT→IT interface pairs at the top and log it; if a later rule is accidentally placed above it, the log shows it immediately. In the dual-firewall model, some organisations choose the two devices from different manufacturers. This makes it harder for the same vulnerability to affect both boundaries, but requires expertise in two different interfaces. The choice is driven not by device brand but by the expected number of rules and frequency of change.
Servers in the DMZ and their roles
Only servers that act as intermediaries go into the DMZ. In this design, systems with a control function (SCADA server, engineering workstation, OT domain controller) are not placed in the DMZ. The layout below is a common example; it is adapted in each plant according to the flows and the risk assessment. Typical components:
- Historian replica: A read-only copy of the data in the OT historian. Reporting, MES/ERP and analytics users connect to it; they do not connect directly to the historian in OT.
- Patch distribution server: In Microsoft environments, a WSUS replica synchronised from the WSUS in IT; the downstream WSUS in OT pulls updates from this server. Approval authority stays with the OT team.
- Antivirus/EDR signature relay: A relay server that receives signature and policy updates from the IT console and passes them to OT agents.
- Remote access intermediary (jump host / RD Gateway): The single point through which all interactive access passes. MFA is mandatory and session recording is on.
- File transfer server: An SFTP-based staging store for PLC projects, recipes and installation files; files are scanned for malware before passing into OT.
- Time source intermediary: An NTP server or GPS-based local source in the DMZ so that OT does not connect directly to the corporate NTP.
A DMZ server sits only on the DMZ segment, with a single network card. A server connected to both the DMZ and OT with two cards (dual-homed) bypasses the firewall completely; this is one of the first things we look for in an inventory scan, because servers with a second card fitted as a “quick fix” are far from rare in the field. Dividing the DMZ into sub-segments by role rather than a single flat VLAN is also valuable: if data servers, remote access and file transfer are in separate segments, a compromised jump host cannot find a direct path to the historian replica.
Historian replication and data flow direction
The historian is the component generating the most traffic in the DMZ, and flow direction is decisive here. The correct structure is for the historian in OT to send data to the replica in the DMZ; the replica does not initiate connections into OT. Many commercial historians offer a server-to-server replication interface for this. Do not guess the port and direction; verify them from the manufacturer’s documentation. For historian selection and time-series databases, the article choosing a production historian offers further guidance.
Things to check in replication:
- The replica must be read-only; tags or notes written by IT users must not flow back to the OT historian.
- There must be buffering (store-and-forward) on the OT side when the connection drops. Otherwise, every maintenance task in the DMZ leaves a hole in the historian.
- Tags to be transferred to the replica should be selected individually; recipe, safety system and process alarm data do not always need to be visible in IT.
For critical processes where one-way flow must be enforced in hardware, a data diode can be considered instead of a firewall; the comparison is in the article data diodes and unidirectional gateways.
Example flow and rule table
The table below is a starting rule set for a typical DMZ. Ports are examples; base your rules on your own product’s defaults and configuration. In each row, the side that initiates the flow is in the “source” column.
| Source | Target | Service | Purpose |
|---|---|---|---|
| OT historian | DMZ historian replica | Vendor-specific TCP | Data replication (push) |
| IT reporting, MES | DMZ historian replica | Vendor-specific TCP or HTTPS | Read |
| DMZ WSUS | IT WSUS | TCP 8530/8531 | Upstream synchronisation |
| OT WSUS (downstream server) | DMZ WSUS | TCP 8530/8531 | Pulling approved updates |
| IT users (VPN, MFA) | DMZ RD Gateway | TCP 443 | Remote access entry |
| DMZ RD Gateway | OT engineering workstation | TCP 3389 | Session to an authorised target |
| OT file client | DMZ SFTP | TCP 22 | Pulling the scanned file |
| OT servers | DMZ NTP | UDP 123 | Time synchronization |
| IT ↔ OT | — | All | Explicit “deny” and log |
The only DMZ→OT row in the table is RDP from the RD Gateway to engineering workstations. Narrow this rule by target: specific workstations, not the whole OT subnet. Define source user–target mappings on the RD Gateway (e.g. an RD Gateway resource authorization policy) and record sessions. WSUS 8530/8531 are the default HTTP/HTTPS ports on Windows Server 2012 and later; if you configured a different port, write the rule accordingly. After opening the rule, always test from outside the DMZ and confirm in the log that it is denied.
Common design mistakes in the DMZ
- “Temporary” pass-through rules: An IT→OT rule opened during commissioning never gets closed. Give each rule an owner and an expiry date, and periodically clean up those that have expired.
- “Passing” a protocol through the DMZ: Carrying OPC Classic (DCOM) or a database connection into OT via a port forward in the DMZ punctures the termination rule. The intermediary application must receive the data and re-serve it.
- A shared management network: If DMZ servers and OT devices can be reached from the same management VLAN in IT, the firewall has been bypassed through a side path.
- Unrestricted internet access from the DMZ: Internet access for DMZ servers should be limited to the necessary update sources only, preferably via the proxy in IT.
- An unlogged boundary: If DMZ firewall logs do not go to central monitoring, nobody sees the attempts caught by the “deny” rule.
- Keeping backups in OT: If DMZ servers are backed up to the OT network, an unnecessarily wide channel is opened between the DMZ and OT. Design the backup flow separately.
Design checklist
- There is no direct allow line between IT and OT in the rule base; the explicit “deny” rule is logged.
- Source, destination, port, initiating side, owner and justification are documented for every flow.
- DMZ servers have a single network card and are not members of the IT domain.
- The DMZ is divided into sub-segments by role (data, remote access, file transfer).
- The historian replica is read-only; there is buffering on the OT side.
- Remote access is protected with MFA, sessions are recorded and targets are restricted per user.
- File transfers cannot reach OT without a scanning step.
- Patch and signature updates are distributed via the DMZ intermediary, with the OT team’s approval.
- Firewall and server logs go to central monitoring; management access is done from a separate network.
- Temporary rules have an expiry date and there is a periodic review schedule.
The DMZ is only one part of IT/OT segmentation; how the separation is set up on the enterprise side is covered in the guide IT/OT segmentation in production facilities. For the architectural framework, NIST’s SP 800-82 Rev. 3 OT security guide is the primary source. If you are going to inventory existing flows and plan a move to a DMZ, our OT cybersecurity and segmentation and IT/OT integration consulting pages show how we carry out this work.
Frequently asked questions
Is a separate firewall mandatory for the OT DMZ?
It is not mandatory. A DMZ can also be built with a single three-legged firewall. But in this model, a single faulty rule opens both boundaries at once; an explicit, logged deny rule between the IT and OT interfaces and regular rule audits are therefore essential. For critical processes and in plants with a separate OT team, dual firewalls provide a cleaner separation of management.
Should the MES server sit in the DMZ or in OT?
The MES components that talk directly to process equipment and are needed for production to continue stay at Level 3, i.e. on the OT side. The reporting or integration layer that interacts with the ERP and enterprise users exchanges data through intermediary servers in the DMZ. The criterion is simple: if production stops when the connection is lost, the component belongs in OT.
Is it right to connect directly from the jump host in the DMZ to a PLC?
It is not. The jump host is used to open a session on an authorised engineering workstation in OT; access to the PLC is done with the engineering software on that workstation. This way, PLC protocols are not carried into the DMZ, and engineering tools stay at a single, controlled point.
How should a cloud-based monitoring service access OT data?
The cloud service does not connect to OT. Data is sent from OT to an intermediary in the DMZ; the cloud receives data only from this intermediary, or over a connection the intermediary initiates outbound. That way, not a single inbound rule from the internet towards OT is needed.