Enterprise networks are segmented to protect data. Industrial control networks must be segmented to protect physical processes: turbines, chemical reactors, water treatment pumps. In these systems a malicious packet can cause a pressure excursion, not just a breach notification. The engineering stakes are different, and so are the constraints.

Operational technology (OT) networks were built for determinism and longevity. A programmable logic controller installed in 2005 may still run its original firmware and speak a protocol that assumes every node on the wire is trustworthy. Patching windows are measured in annual maintenance outages, not weekly cycles. Availability outranks confidentiality, which inverts the priority order most security architects learned on IT networks.

Segmentation is the primary control available in this environment. It cannot fix an insecure protocol or an unpatchable controller. It can limit who reaches those systems and through which paths. The challenge is drawing those boundaries without breaking the control loops, historian feeds and vendor support sessions the plant depends on. Three architectural patterns define how that balance is struck.

The IT/OT Boundary: Designing the Industrial DMZ

The Purdue Enterprise Reference Architecture, later aligned with ISA-95, organises industrial networks into levels. Field devices sit at Level 0, controllers at Level 1, and supervisory HMIs and SCADA servers at Level 2. Site operations and historians occupy Level 3, and enterprise systems sit at Levels 4 and 5. The critical engineering decision lies between Levels 3 and 4, at what practitioners call Level 3.5: the industrial demilitarised zone, or IDMZ.

The governing rule of a well-designed IDMZ is that no session traverses it end to end. Traffic from the enterprise terminates on a server inside the DMZ, and a separate session runs from that server toward the control network. A replicated historian is the standard example. The plant historian pushes data to a mirror in the IDMZ, and business users query the mirror. Enterprise clients never connect to anything below Level 3.5. Firewalls on both sides enforce this, ideally from different vendors or at least with independently managed rule sets.

Where the risk justifies it, data diodes replace the outbound firewall entirely. A unidirectional gateway enforces one-way flow in hardware, typically a fibre link with a transmitter on one side and only a receiver on the other. TCP requires acknowledgements, so diode vendors run proxy software on each side that terminates the connection and rebuilds it. This lets historian databases and file transfers replicate with no return path. Nuclear facilities and some grid operators require this pattern.

Diodes cannot carry inbound access, so that is handled through jump servers. An engineer or vendor authenticates to a hardened host in the IDMZ using multi-factor credentials, then opens a brokered session to the target system. The jump host becomes a choke point for logging, session recording and time-bound approvals. Its value depends entirely on discipline. A jump server with persistent credentials and broad reachability simply moves the vulnerability somewhere else.

Takeaway

A DMZ is defined less by its firewalls than by the rule that no connection crosses it intact. If any session flows straight from the enterprise to the control network, the zone exists only on the diagram.

Zones and Conduits: The ISA/IEC 62443 Model

The Purdue model describes a hierarchy. It says little about how to subdivide each level or how much protection a given segment needs. ISA/IEC 62443 fills that gap with two abstractions. A zone is a group of assets that share common security requirements. A conduit is a logical group of communication channels connecting zones, and it carries its own security requirements.

Zones are defined by risk assessment, as specified in IEC 62443-3-2, not by network topology. A safety instrumented system belongs in its own zone even if it shares a cabinet with the basic process control system, because the consequences of compromising it are different. Each zone receives a target security level, or SL-T. The scale runs from SL 1, protection against casual or coincidental violation, to SL 4, protection against sophisticated, well-resourced attackers with IACS-specific skills.

In practice the model separates three kinds of security level. SL-T is what the risk assessment demands. SL-C is the capability a system or component can deliver, measured against requirements in 62443-3-3 and 62443-4-2. SL-A is what the deployed system actually achieves once configured. The gap between target and achieved levels becomes the engineering work list. That list can include compensating controls, upgraded components, or a conduit with stricter enforcement to protect a zone whose devices cannot meet the target on their own.

Conduits are where the model becomes concrete. Each conduit records which zones it connects, which protocols it carries, which direction traffic flows, and which enforcement points sit on it, such as firewalls, VPN concentrators or diodes. This turns segmentation from a VLAN plan into an auditable specification. Suppose a new requirement appears, such as a cloud analytics feed from a packaging line. The question becomes which conduit the flow belongs to and whether that conduit's security level permits it, not simply which port to open.

Takeaway

Zones answer what needs protecting and how much. Conduits answer how those zones are allowed to talk, and designing them explicitly keeps segmentation from eroding one exception at a time.

Protocol Considerations: When the Wire Assumes Trust

Most industrial protocols predate the idea that a control network might contain an adversary. Modbus TCP, listening on port 502, has no authentication. Any client that can reach a device can read registers or write coils. DNP3, EtherNet/IP and PROFINET were also designed for closed environments where being able to reach a device implied permission to use it. Segmentation therefore carries a burden it rarely bears in IT: it is often the only access control the protocol has.

Some protocols actively resist segmentation. PROFINET real-time traffic runs directly over Ethernet at Layer 2 and cannot be routed, so controllers and their I/O must share a broadcast domain. EtherNet/IP implicit messaging uses UDP port 2222 and is frequently multicast, which complicates router and firewall configuration. OPC Classic relies on DCOM with dynamically allocated ports. That makes clean firewall rules nearly impossible without tunnelling or migrating to OPC UA, which uses a single configurable port and built-in security.

Port-based filtering is also too coarse to mean much. Allowing TCP 502 allows every Modbus function, from reading holding registers (function code 3) to writing multiple registers (function code 16), plus diagnostic functions that can force a device offline. Industrial deep packet inspection firewalls parse these protocols and enforce rules by function code, register range or CIP service. They can let a historian read but never write, or allow PLC program downloads only from a designated engineering workstation.

Inline inspection brings its own trade-offs. Every enforcement point adds latency and jitter, and a control loop with tight timing may not tolerate either. Failure modes also need explicit decisions. An inline device that fails closed may halt production, while one that fails open quietly removes protection. For this reason many sites combine inline enforcement at conduit boundaries with passive monitoring through network TAPs or SPAN ports inside zones. This gives protocol-level visibility without placing anything in the real-time path.

Takeaway

When a protocol has no authentication, the network path becomes its permission model. What a packet asks a device to do matters more than which port it arrives on.

Effective OT segmentation starts from the process, not the topology. Identify the consequences of compromise, group assets into zones accordingly, and assign target security levels before drawing a single VLAN.

Then enforce a small set of non-negotiable principles. No session crosses the IDMZ end to end. Every conduit is documented with its protocols, direction and enforcement points. Remote access flows through a monitored jump host with time-limited approvals. Industrial traffic is filtered by function, not just port, wherever latency budgets allow.

Legacy devices will outlive many of today's controls, and some protocols will never authenticate their peers. The architecture around them has to compensate: deliberately, auditably, and without treating availability as an afterthought.