IMD SolutionComputing LibraryShop main store

How-to

Segmenting edge platform networks for data, BMC and OS traffic

OT cybersecurity Published 7 min read

Short answer

Three interfaces are the right starting point, but they isolate nothing by themselves. Put data on the industrial segment, BMC and OS management on a restricted management segment reached only from named administrative systems, feed updates through approved destinations instead of the open Internet, and disable host forwarding so the edge computer cannot bridge zones. IEC 62443-3-2 frames these boundaries as zones and conduits; every crossing needs an explicit rule.

On this page
  1. What three interfaces fix, and what they do not
  2. One trust boundary per traffic role
  3. Map the interfaces and dependencies first
  4. Build the segments and the controlled paths
  5. Worked example: six transit paths across three interfaces
  6. Prove the boundary holds
  7. When the platform is already wired broadly
  8. Questions asked

Three interfaces — data, BMC and operating-system management — is the right starting pattern for an industrial edge platform, and the wiring claims that usually ride with it are where the design goes wrong. The OS interface does not take direct Internet access, and neither BMC nor OS management belongs on the general corporate LAN. Put data on the industrial segment, put BMC and OS management on a restricted management segment reached only from named administrative systems, and feed updates from approved destinations. IEC 62443-3-2 gives the frame: group assets into zones that share a security requirement, and let traffic cross a zone boundary only through a conduit you have declared and control. The steps below take you from an interface map to an acceptance test.

What three interfaces fix, and what they do not #

Separate adapters divide traffic; they do not remove the host's ability to communicate on every network it is attached to. Four common choices, and the correction each one needs:

Common choice Why it falls short The correction
One adapter per role: data, BMC, OS The host can still route, bridge or proxy between its networks. Disable forwarding, remove bridges, enforce per-interface firewall policy.
OS interface directly on the Internet for updates Broad egress with no approval, staging or rollback. Update from an internal mirror or a rule naming specific approved destinations.
BMC kept local only Locality names no authorized administrator. A restricted management segment with an allow-list of administrative sources.
BMC and OS management on the corporate LAN Every corporate endpoint becomes a possible peer of privileged services. Reach management only from designated administrative systems.

The design to commit to is four sentences: data on the industrial segment, BMC and OS management on a restricted management segment, updates through an approved path, host forwarding off. Everything below is those four statements made enforceable.

One trust boundary per traffic role #

IEC 62443-3-2 defines the segmentation model used across its system-level requirements: a zone groups assets that share a security requirement, and a conduit is a declared, controlled communication path that crosses a zone boundary. CISA's guidance for edge devices points the same way — keep edge devices isolated from critical systems and grant least-privilege access. For a multi-homed edge computer, each interface is the edge of a zone, not just a cable:

Traffic role Intended communications Boundary to enforce
Data The enumerated industrial devices and the applications consuming their data Peers, protocols and connection direction limited to application needs.
BMC Authorized administrators and management services Separate from production data and general corporate endpoints.
OS management Administration, monitoring, approved updates, required upstream services Only documented administrative and service flows permitted.

OS management does not require unrestricted Internet access. Updates can come from an internal mirror or a firewall rule naming specific destinations, and upstream application traffic can use a restricted gateway. Outbound-initiated connections reduce inbound exposure, but judge the service behind the connection, not its direction: an outbound tunnel can still carry remote commands or move data out.

The BMC deserves separate treatment because a management controller runs independently of the host OS and keeps its access while the host is down. Treat corporate-network reachability as a firewall policy decision — a list of authorized administrative systems — never as permission for every corporate endpoint.

Segmentation is also one layer among several; it works alongside the layered OT security controls that protect a network path, not instead of them.

Map the interfaces and dependencies first #

Build the topology from a communication inventory, not from network labels like "local" and "corporate" — those are too broad to configure a switch, a host firewall or a route with.

  1. Record each device's physical adapters, VLAN memberships, addresses, management controller and supported remote-management method. Mark the devices that have no BMC; they are a different hardware class and take a different map.
  2. List every data source and destination with its application protocol, which side initiates the connection, and whether discovery or broadcast traffic is needed.
  3. List administrative dependencies: name resolution, time synchronization, identity, monitoring and software distribution where used.
  4. Inspect the host routing table, the default route, forwarding settings, bridges and listening services — and container or virtualization networking if the applications use it.
  5. Record which management services listen on which interface. A service bound to all addresses is reachable from the data network even though administrators normally use another interface.
  6. Write the permitted source, destination, service and direction for every required flow. This list becomes the firewall rules and the acceptance test.

Where device discovery depends on a shared broadcast domain, test it before segmenting, because routing does not carry local broadcasts across a boundary. For the flows an edge application legitimately opens, the peer, protocol and direction enumeration in bridging OPC UA PLC data to MQTT with an edge PC is a concrete model to copy.

Build the segments and the controlled paths #

  1. Back up the host network configuration, the switch configuration and the access rules, and establish an approved console or recovery path before touching management connectivity.
  2. Connect the data interface to the industrial segment and permit only the device communications the application needs — the same peer-level discipline as connecting a PLC to the plant network without exposing drives. Disable unused services on that interface.
  3. Put BMC interfaces on the restricted management segment, with access from designated administrative systems only.
  4. Put OS administration on its management segment and restrict administrative listeners and host firewall rules to authorized sources.
  5. Configure the default route and any specific routes deliberately, so data destinations stay on their intended path and no interface becomes transit.
  6. Provide upstream connectivity through controlled firewall or gateway rules: software distribution and application egress to approved services, nothing broader.
  7. Check that host bridges, routing, proxies and container networking do not recreate a path the external firewall blocks.

Schedule the change with operations, because removing a route that carries an application dependency can stop data collection. On switching you already own, segregating machine traffic from the plant network shows how far VLANs and access lists on existing hardware carry these same boundaries.

On a small single-port device, data and OS management can share the adapter using tagged VLANs and per-role policy. What cannot be shared is the trust boundary.

Worked example: six transit paths across three interfaces #

Inputs: the proposed platform topology starts from three interfaces — data, BMC and OS management — which is the interface count this design assumes. The engineering fact behind the count: a multi-homed host with forwarding enabled, a bridge configured, or a proxy running can carry traffic between any two of its interfaces, in either direction.

Rule in words: the number of ordered interface pairs, which is the number of possible transit paths, equals the interface count multiplied by one less than itself.

Arithmetic: 3 × (3 − 1) = 3 × 2 = 6.

The six paths: data to BMC, BMC to data, data to OS, OS to data, BMC to OS, OS to BMC.

Result: six directional transit paths to account for on one edge computer.

Decision: every one of the six is either blocked by policy or declared as a conduit with a documented reason. The threshold is zero unexamined pairs, because an unexamined pair is an undeclared conduit. A fourth interface, such as a cellular uplink for remote access, raises the count to 4 × 3 = 12 and doubles the review; add interfaces only when the role justifies that.

Prove the boundary holds #

The flow list from step 6 of the inventory is the acceptance test. Run it after cutover and after every OS or firmware update.

  1. From each segment, attempt each of the six directional interface pairs and every disallowed service, and expect every attempt to fail.
  2. Attempt each documented production and management flow, and expect each to pass.
  3. Confirm forwarding is disabled, no bridge spans role interfaces, and each administrative listener answers only on its management address.
  4. Record the results with the configuration baseline so the next change starts from a known state.

When the platform is already wired broadly #

Contain first, on a schedule operations agrees to: restrict unnecessary management access and outbound traffic while the validated production flows keep running. Then repair permanently by documenting and enforcing the allowed communications. A temporary cable arrangement is a staging state, not a security boundary; the boundary is the enforced rule set you can test.

Questions asked

Is a separate network adapter per role enough to isolate the networks?

No. An adapter divides cables, not trust. With IP forwarding enabled, a bridge configured, or a service listening on all addresses, the host still carries traffic between roles. Isolation comes from disabled forwarding, no bridges, listeners bound per interface, and a default-deny host firewall with per-role allow rules, reinforced by access lists on the segment's switch.

Should the OS interface get direct Internet access for updates?

No. Give it egress only to approved update destinations, such as an internal mirror or a firewall rule naming the vendor hosts, with connections initiated outward. Direct access adds exposure without adding staging, approval or rollback. Judge the service behind an outbound connection too: a tunnel can still carry remote commands or move data out, whatever its direction.

Where does a BMC belong when there is no out-of-band management network?

On a restricted management VLAN with an access list naming the administrative workstations or jump host, never on the general corporate LAN. A management controller runs independently of the host OS and often provides console, media and power control, so its peer list should be a set of systems, not a network. If the switching cannot enforce that yet, keep the controller physically local until it can.

Can data and OS management share one adapter on a small edge device?

Yes. Tag the single port for two VLANs and write firewall policy per sub-interface, so each role keeps its own zone and rules. What must not be shared is the trust boundary: one flat policy across both roles puts administration traffic and plant data in the same conduit.

How do I prove the segmentation actually blocks transit?

Attempt the six directional interface pairs from the worked example from each segment and expect every one to fail; then attempt each documented production and management flow and expect each to pass. Record the result as the acceptance test, and re-run it after OS or firmware updates, which can re-enable forwarding or rebind a listener to all addresses.

Keep reading

  • Comparison

    Why layered OT security controls protect a network path

    No single control secures an OT request path. Network reachability, session protection, user identity and host hardening each block a different failure, so commissioning verifies them in that order: route first, then TLS and certificates, then login and role permissions, then the hardened baseline. A host that answers a connection proves none of the other three.

    OT cybersecurity

  • Comparison

    Connect a PLC to the plant network without exposing drives

    Keep the PLCs, HMIs and drives on one dedicated machine subnet and connect it upward through exactly one controlled boundary. A routed firewall, a second Ethernet interface such as the CP 1243-1, or a 1:1 NAT switch of the S615 class all do the job; the deciding factors are which CPU you have and which side initiates each connection.

    OT cybersecurity

  • Checklist

    Verify SCADA data and command paths before assuming an attack

    Most wrong or frozen SCADA values are tag, binding or driver faults, not attacks. Compare the display, the SCADA tag with its quality and timestamp, and the controller value: agreement clears the data path, and only unexplained writes or unauthorized reachability justify a cyber investigation.

    OT cybersecurity