Why Segmentation Stalls in GMP Plants

Network segmentation means dividing one network into smaller parts and controlling what traffic may pass between them. In industrial settings the usual language comes from the ISA/IEC 62443 series, developed by ISA’s ISA99 committee (established in 2002) and adopted by the IEC as international standards.7 A zone is a group of assets with common security requirements. A conduit is the controlled path between zones. FDA’s 2025 white paper on manufacturing OT uses the same terms and gives a pharma example: separate fill-finish lines placed in separate zones, so that a problem on one line does not spread to the others.1

None of this is controversial in a security meeting. The trouble starts when the design reaches the plant. Consider what a typical GMP site network carries on a normal shift:

  • Weights and dispensed quantities flowing from scales and dispensing stations into a manufacturing execution system (MES) or electronic batch record.
  • Critical process parameters flowing from controllers into a historian, and from the historian into reports used for batch review.
  • Audit trail entries written by instruments and supervisory systems to central databases.
  • Time signals that every one of those systems uses to stamp its records.
  • Backups, antivirus updates, user authentication, and vendor support sessions.

Each of these is a data flow. On a flat network nobody had to write them down, because everything could reach everything. Once a firewall is placed between two groups of systems, every flow that is not explicitly allowed stops. If one of those flows carried a GMP record, the plant now has a data integrity problem, and possibly a product quality problem.

The Network Is Part of the Validated State

Quality teams sometimes treat the network as something outside the scope of validation. EU GMP Annex 11 does not. Its principle section says, “The application should be validated; IT infrastructure should be qualified.”8 Annex 11 defines IT infrastructure as the hardware and software, such as networking software and operating systems, that makes it possible for the application to function. ISPE makes the same point in its GAMP good practice guide on IT infrastructure: “The validated status of GxP applications that are dependent upon an underlying IT Infrastructure can be compromised if the IT Infrastructure is not maintained in a demonstrable state of control and regulatory compliance.”9

So a segmentation project changes qualified infrastructure that validated applications depend on. Annex 11 section 10 requires that “Any changes to a computerised system including system configurations should only be made in a controlled manner in accordance with a defined procedure.”8 A firewall rule set is a system configuration. This is not a reason to avoid segmentation. It is a reason to plan it the way quality plans any other change to validated systems.

Four Patterns Behind Stalled Projects

In our experience, segmentation projects at GMP sites stall for a small number of repeatable reasons:

  1. The design comes before the data flows. A network architect draws zones from an asset list and a reference model. The first time quality asks which validated flows cross each new boundary, nobody can answer, and the change control board will not approve a change whose impact is unknown.
  2. Validation documents describe applications, not paths. A system’s validation package usually describes the interfaces in functional terms (“the MES receives weights from the scale”). It rarely records the IP addresses, ports, and protocols those interfaces use, or what the system does when the path is blocked.
  3. One large change for the whole site. A single change record that moves dozens of systems at once is too big to assess, too big to test in one window, and too big to roll back cleanly.
  4. Time and audit trails are an afterthought. Teams test that the application still works but do not check that its clock still synchronizes or that audit trail entries still reach their database. These problems often appear weeks later, during a batch review or an inspection.

The rest of this article addresses each of these in turn. For a broader view of plant cybersecurity (the threat picture, vendor remote access, legacy equipment, and pre-assessed categories for routine security changes), see our earlier article on preemptive cybersecurity for the connected pharma plant. Here we stay focused on one job: getting a segmentation design into a running GMP plant without breaking what validation depends on.

Start With a GMP Data Flow Register

FDA’s white paper is direct about the order of work. It says the zone and conduit approach “requires an accurate inventory of all connected assets and a clear understanding of their roles and communication needs within the industrial process.”1 NIST SP 800-82 Revision 3, the main US government guide to OT security, says organizations typically use their mapped data flows to identify required communications, then configure those requirements into the network devices that enforce segmentation.2

Most sites read this as “build an asset inventory,” and that is a good first step. But an asset inventory answers the question “what is on the network.” Segmentation needs the answer to a different question: “what has to talk to what, and why.” For a GMP plant, it also needs a third answer: “which of those conversations carry or support a GMP record.”

What an Asset Inventory Leaves Out

The August 2025 joint guidance from CISA and partner agencies on OT asset inventories is a good model for the inventory itself. It sets out a process of defining scope, identifying assets and collecting attributes, building a taxonomy, managing the data, and managing the asset life cycle. Among the attributes it lists as high priority are the active or supported communication protocols and the asset’s criticality.3 On conduits, it says, “To determine Conduits, organizations should map communication pathways between zones incorporating data flow analysis, protocol identification, and Layer2/Layer3 network details.”3

That is the gap. An inventory row describes one device. A conduit is defined by the conversations between devices. A GMP site needs both, and it needs the conversations described in terms quality can assess.

Fields for Each GMP Data Flow

We recommend a separate register with one row per data flow, linked to the asset inventory by asset identifier. A practical set of fields:

FieldWhat to RecordWhy Quality Needs It
Flow ID and nameA unique identifier and a plain description (“Line 2 checkweigher to MES, reject counts”)Lets change records, test scripts, and firewall rules refer to the same flow
Source and destinationAsset IDs, hostnames, IP addresses, and the zone each will belong to after segmentationShows whether the flow will cross a new boundary
Protocol, ports, directionFor example OPC UA on a fixed port, SQL, SMB file share, syslog, NTP; who starts the connectionBecomes the firewall rule; direction matters for data diode decisions
GMP record carriedThe record or data element: batch record entry, critical process parameter, audit trail, alarm, electronic signatureDrives the GMP impact tier
Behavior when blockedBuffers and resends, raises an alarm, stops the process, or drops data without warningDecides how dangerous a mistake in this rule would be
Time sourceWhich clock the source and destination useProtects the timestamps in both records
ScheduleContinuous, per batch, daily, monthly, on demandTells the test team how long to watch after a cutover
Validation referenceThe validated system and the document that describes this interfaceIdentifies which validation documents need updating
Verification evidenceWhat objective evidence shows the flow works (a record in the destination that matches the source)Becomes the acceptance criterion in the test
OwnerThe system owner and the technical contactTells the change board who must approve and who must test

This register is the single most useful document in the project. Network engineers turn it into rule sets. Quality uses it to scope impact assessments. Validation uses it to write test scripts. And after go-live, it becomes the reference for periodic review of the firewall rules.

Finding the Flows Nobody Wrote Down

Interviews and validation documents will find most flows. They will not find all of them. FDA’s white paper describes two problems that make discovery hard: equipment that contains embedded control devices from other manufacturers, and connections that the equipment maker does not disclose to the end user.1 The same paper gives an example of third-party firmware that tries to communicate externally, a flow that may not appear on an initial mapping but would be discovered during security scans.1

NIST offers a practical method for this. When it is hard to determine which traffic is needed, it says, “organizations might consider temporarily allowing and recording all communication between the network segments.” NIST notes this can also reveal previously unknown or undocumented communication that needs review.2 In practice that means placing passive monitoring or a firewall in a log-only mode at the future boundary, and recording what crosses it before any rule blocks anything.

Two cautions for GMP sites. First, the observation period has to cover the full operating cycle, not a convenient two weeks. Monthly archive jobs, quarterly report runs, annual calibration uploads, and campaign changeovers all produce flows that a short sample will miss. Second, every flow discovered this way needs an owner and a decision. Some will be essential and undocumented, which is itself a finding for the validation file. Others will be unneeded, such as a vendor’s update check that nobody approved, and can be blocked on purpose.

Sort Flows by GMP Impact

Not every flow needs the same rigor. A simple three-tier scheme keeps the later work proportionate:

Tier 1

Record-Forming Flows

The flow creates or changes a GMP record: weights to the batch record, critical parameters to the historian used for release, audit trail entries, electronic signatures, time signals for any of these.

Tier 2

GMP-Supporting Flows

The flow does not create a record but supports a GMP activity: user authentication, backup and restore, alarm notification, recipe download, antivirus and patch distribution to validated hosts.

Tier 3

Non-GMP Flows

Business or convenience traffic with no GMP impact: energy dashboards, office printing, general internet access from engineering laptops, building systems with no GMP role.

Unknown

Unexplained Flows

Discovered traffic with no owner yet. Treat as Tier 1 until someone shows otherwise, and track each one to a decision before its zone is cut over.

The tier drives everything downstream: how much testing a rule needs, whether a cutover can happen during a production run, and how quickly a rollback must be possible.

Design Zones Around Validated System Boundaries

With the register in hand, zone design becomes a matter of grouping. NIST lists several bases for grouping devices, including management authority, level of trust, functional criticality, data flow, and location, and suggests organizing segmentation with an industry model such as the Purdue model, ISA-95 levels, the three-tier IIoT architecture, or a combination.2 The Purdue model numbers levels from field devices (Level 0) through control (Levels 1 and 2), site operations (Level 3), and enterprise systems (Level 4 and above), usually with a demilitarized zone (DMZ) between site operations and the enterprise.

Rules Between Levels

NIST is specific about how traffic should move between levels. Rule sets should only permit connections between adjacent levels, tiers, or zones, and in a Purdue architecture Level 4 devices should not communicate directly with devices at Levels 2, 1, or 0.2 NIST also flags a common weakness: outbound traffic from the control network. It says, “Organizations should consider making outbound rules as stringent as inbound rules to reduce these risks.”2

The adjacent-level rule has a direct effect on GMP data flows. On many older sites, an enterprise reporting tool or a quality system queries a controller or a line historian directly. Under the new design, that flow has to be rerouted through a Level 3 system or a DMZ replica. Rerouting changes the path of a GMP record, which means it is not a firewall change alone. It may need an interface change in the validated system, with its own testing.

Match Zone Boundaries to Validation Boundaries

The most useful design rule we know for GMP sites is simple: where you can, keep each validated system inside one zone. If a chromatography data system’s servers, acquisition controllers, and client workstations are all in the same zone, the traffic among them never crosses a firewall, and the system’s internal behavior does not change when segmentation goes in. Only its external interfaces cross conduits, and those are the ones you test.

When a validated system is split across zones, every internal connection becomes a conduit. That multiplies the rules, the tests, and the ways the system can fail. Sometimes a split is right, for example when client workstations are on the office network and servers belong in the plant. But it should be a choice made with the system owner, not a side effect of drawing zones by device type.

A few other boundary questions come up at almost every site:

  • Line or suite zones. FDA’s fill-finish example suggests one zone per line.1 For a biologics site, a zone per suite or per unit operation group often matches how batches are run and released.
  • Laboratory systems. QC instruments and their data systems often belong in their own zone, separate from production, with conduits to LIMS and the data archive.
  • Safety systems. NIST says that while VLANs can be a less expensive option, organizations should consider physically separate switches for high-criticality devices such as those that support safety systems.2
  • Engineering workstations. These can program controllers, so they need tight placement and tight conduits. They are also often the machines that run monthly or ad hoc jobs nobody remembers until they fail.

Decide Where Shared Services Go

Shared services touch almost every zone, so their placement decides a large share of the rule set. Plan these explicitly:

  • Time servers. Covered in detail later. Every validated system that stamps records depends on them.
  • Directory and authentication services. Many plants run a separate OT domain rather than extending the corporate one into production zones. Either way, if authentication stops working, operators cannot sign records.
  • Backup. Backup traffic is often the largest flow in an OT network and the one most likely to be forgotten in testing, because it runs at night.
  • Patch and antivirus distribution. Usually staged in the DMZ, with conduits down to each zone.
  • Historian replication. A common design keeps a site historian in Level 3 and a replica in the DMZ for enterprise users, so nothing from Level 4 reaches the plant historian directly.

Firewalls or Data Diodes

Both NIST and the NCSC-led secure connectivity principles discuss unidirectional gateways, also called data diodes, which allow data to travel in one direction only.24 They give stronger protection than a firewall for flows that only need to leave the plant, such as historian replication to the enterprise.

For GMP purposes the question is whether the flow can live without a return path. Many industrial and database protocols expect an acknowledgment. If a flow depends on two-way traffic and is pushed through a one-way device, the vendor’s replication software has to handle that, and the design has to be tested to show the destination record is complete. The data flow register’s “direction” field is where this decision starts.

Run One Risk Assessment for Security and Quality

ISA/IEC 62443-3-2 sets the requirements for the security risk assessment behind zone design. The IEC lists them as defining a system under consideration, partitioning it into zones and conduits, assessing risk for each zone and conduit, establishing a target security level for each, and documenting the security requirements.5 The ISA Global Cybersecurity Alliance white paper on the standard describes the same work as a sequence: identify the system under consideration, perform an initial cyber risk assessment, partition into zones and conduits, perform a detailed risk assessment, and document the updated requirements.6

A GMP site already does quality risk assessment for its validated systems. Running two separate assessments on the same zones invites contradictions. Instead, add a short set of GMP questions to each zone and conduit in the 62443 assessment:

  • Which Tier 1 and Tier 2 flows cross this conduit?
  • If this conduit fails closed (blocks all traffic), what happens to each flow: buffer, alarm, stop, or loss of data?
  • How long can each flow be interrupted before a record is lost or a batch is affected?
  • Is there a manual or alternative way to continue the GMP activity while the conduit is down?
  • Which validated systems’ documents describe the flows through this conduit?

The answers feed the security level decision, the change control sequence, and the rollback plan. They also give quality reviewers a record of how the design accounted for product and data risk, which is the question an inspector is most likely to ask.

Change Control and Impact Assessment, Zone by Zone

FDA’s white paper notes that a change control board can help maintain visibility and raise shared concerns when projects use shared enterprise resources, by reviewing each proposed change before it is made.1 In a GMP plant this is not optional, and the way the change records are structured decides whether the project moves.

One Program Change, Many Wave Changes

We recommend two levels of change record:

  • A program-level change that approves the target architecture, the data flow register as the governing reference, the GMP tiering, the test strategy, and the standard rollback approach. Quality approves the method once.
  • A wave-level change for each zone or small group of zones, which references the program change and covers only the systems, flows, and rules in that wave. Each wave has its own impact assessment, test protocol, cutover window, and rollback plan.

This structure keeps each wave small enough to assess properly and to roll back on its own. It also means that a problem in one wave does not hold up approval of the design for the whole site.

What a Zone Impact Assessment Covers

A useful impact assessment for a segmentation wave goes well beyond “the network will change.” It should address, for each system in the wave:

AreaQuestions to Answer
AddressingWill the system’s IP address, subnet, gateway, or hostname change? Is the address hard-coded anywhere: controller logic, HMI projects, recipes, configuration files, license files, scripts?
Data flowsWhich register flows cross a new boundary? Is each one covered by a rule? Has each been assigned a test?
TimeWill the system’s time source change or move behind a new firewall? Is NTP or PTP allowed to it?
AuthenticationCan users still log in and sign records? Does the system reach its domain controller, and does it fail safely if it cannot?
Audit trailWhere does the audit trail get written? Does that path cross a new boundary? What happens to entries if the path is blocked?
Backup and restoreDoes the backup job still reach its target? Has a restore been tested through the new path?
Remote supportHow does the vendor reach the system after the change? Is that path documented and approved?
DocumentationWhich validated documents need updating: network diagrams, configuration specifications, interface specifications, system descriptions?
Production scheduleWhich batches will be running during the window? Can the wave wait for planned downtime?

Hidden Dependencies That Break on Cutover Day

Some failures repeat often enough that they deserve their own checklist. Most come from moving a system to a new subnet rather than from the firewall rules themselves:

  • Hard-coded addresses. A controller program or HMI project that refers to a peer by IP address will stop talking to it after re-addressing, even if every firewall rule is correct.
  • Licenses tied to network identity. Some industrial software ties its license to a host identifier or address. Changing the network identity can stop the software from starting.
  • Discovery and broadcast traffic. Some industrial protocols find peers by broadcast or multicast, which does not cross a routed boundary unless the design allows for it.
  • Dynamic ports. Older protocols that negotiate ports at run time are hard to allow through a firewall without opening wide ranges. Plan to replace them, wrap them, or keep both ends in one zone.
  • Name resolution. If systems find each other by hostname, the DNS or host file entries must be right in the new zone.

Each of these can be found before the window by reviewing configurations with the system owner and the vendor. None of them is found by reviewing the firewall rule set alone.

Sequence Waves by Risk and Schedule

Wave order matters. A sequence that works at many sites:

1

Build the Boundary in Monitor Mode

Install the firewalls or gateways that will separate enterprise and plant, with logging on and nothing blocked. Use the logs to complete and correct the data flow register.

2

Stand Up Shared Services

Build the DMZ, the OT time source, the patch staging server, and any historian replica. Move enterprise consumers to the replica and prove the replica’s data matches the source before cutting any direct path.

3

Cut Over Low-Risk Zones First

Start with zones that carry mostly Tier 3 flows, such as building systems with no GMP role or utility monitoring. The team learns the cutover routine where a mistake has little impact.

4

Cut Over GMP Zones in Planned Downtime

Move production and laboratory zones one at a time, during scheduled shutdowns or between campaigns, with the system owners on site and test scripts ready.

5

Close Temporary Rules

After each zone has run cleanly through a full cycle, remove any temporary or broad rules that were kept for safety during the move, under the same change record.

The monitor-first step deserves emphasis. It turns segmentation from a single high-risk event into a period of observation followed by a series of small changes, each based on evidence of what the network carries.

Testing That Proves GMP Data Flows Still Work

The test question for segmentation is narrower than for a new system. You are not proving the application works. That was proven at validation. You are proving that every flow the application depends on still delivers complete and correct records through the new path, and that the paths you meant to close are closed.

Test the Rule Set Before It Reaches Production

NIST SP 800-41, the NIST guide on firewalls and firewall policy, says new firewalls should be tested on a test network without connectivity to production that replicates the production network as faithfully as possible. Among the aspects to evaluate, it lists connectivity, application compatibility, logging, and performance, and on the rule set it says, “Verification of the ruleset should include both reviewing it manually and testing whether the rules work as expected.”10

Most plants cannot build a full copy of a production line. They can usually do three things. They can review each proposed rule against the data flow register, line by line, with the system owner. They can test rules for systems that have a non-production instance, such as a test MES or a QC data system validation environment. And they can run the new rule set on the production firewall in log-only mode for a period, recording which traffic the new rules would have blocked, before switching enforcement on. That last step finds missing rules with no impact on production.

Positive Tests and Negative Tests

Each wave’s test protocol should include two kinds of test:

  • Positive tests prove that each allowed flow works. One test per Tier 1 and Tier 2 flow in the wave, using the verification evidence recorded in the register.
  • Negative tests prove that paths meant to be closed are closed. For example, an attempt from an enterprise workstation to reach a Level 1 controller should fail, and the firewall should log the attempt.

Negative tests are often skipped in GMP projects because they do not relate to an application requirement. They should not be. A segmentation change that leaves an unintended path open has not achieved its purpose, and the negative test is the objective evidence that it did.

Evidence That Shows a Flow Is Complete

“The interface is working” is a weak test result. A strong one compares records at both ends. The PIC/S data integrity guidance says interfaces should be assessed during validation to ensure the correct and complete transfer of data, and lists checks such as secure transfer, encryption, and checksums.11 The same thinking applies after a network change. Good evidence for a Tier 1 flow looks like this:

Flow TypeWeak EvidenceStrong Evidence
Instrument to batch record“MES shows connected”A test weighing appears in the batch record with the same value, unit, and timestamp as the instrument’s own record
Controller to historian“Trend is updating”A trend for the cutover window with no gap, and a count of stored values matching the expected scan rate
Audit trail to central database“No errors in the log”A deliberate test action on the source system, found in the central audit trail with the correct user, action, and time
File transfer to archive“Job completed”File counts and checksums at source and destination match for the test run
Historian replica to enterprise“Users can see data”A sample of values compared between site historian and replica for the same tags and time range

Much of this evidence already exists in normal operation. The point is to name it in advance as the acceptance criterion for each flow, so testers collect it on purpose rather than reconstructing it afterward.

Watch Through a Full Cycle

A flow that runs once a month will not be tested by a cutover on a Tuesday. The register’s schedule field tells the team which flows need a watch period after go-live and how long it must last. For each wave, set a watch period that covers at least one run of every scheduled flow in the zone, and do not close the wave’s change record until each has been seen to work. During that period, keep the firewall’s deny logs under daily review. A blocked connection from a validated system is either a missing rule or an unexplained flow, and both need a decision.

A wave is done when: every Tier 1 and Tier 2 flow in the zone has passed a positive test with record-level evidence; negative tests show intended blocks are in place; clocks on every system in the zone are within the site’s tolerance; audit trail entries from each system have reached their destination; every scheduled flow has run at least once; the deny log has no unexplained entries from validated systems; and the validated documents have been updated.

Rollback Plans and Isolation Plans

Every wave needs a rollback plan written before the window opens. Network changes are usually easy to reverse in principle, which is exactly why teams skip the planning. Under pressure in the middle of a shutdown, “we will just put it back” turns into an improvised change on a validated system.

Decide the Triggers in Advance

A rollback plan should state:

  • The decision point. A fixed time in the window by which all Tier 1 tests must pass. If they have not, the team rolls back rather than troubleshooting into the restart of production.
  • The triggers. Any Tier 1 flow failing its test, any loss of time synchronization on a record-forming system, any loss of audit trail delivery, or any failure that the system owner judges a risk to product or data.
  • Who decides. Named people from engineering, quality, and the system owner, with authority to call it.
  • The exact steps. The saved prior configuration of each firewall and switch, the original addressing for each system, cable and port labels where physical moves are involved, and the order in which to restore them.
  • The verification after rollback. The same Tier 1 tests, run again, to prove the old path works. Rolling back is a change too, and it needs evidence.

A broad “allow all” rule is not a rollback. When a cutover struggles, the tempting fix is to add a temporary rule that permits everything between two zones and sort it out later. That leaves the plant in a state nobody assessed or tested, with neither the old protection nor the new design. If the plan needs a fallback, the fallback should be the previous, documented configuration. Any temporary broad rule should be treated as a deviation, recorded, time-limited, and closed under change control.

Data Produced During a Failed Cutover

The harder rollback question is about data, not configuration. If a flow was blocked for forty minutes before the team noticed, what happened to the records produced in that time? The answer depends on the “behavior when blocked” field in the register:

  • Buffered and resent. Many historians and some MES interfaces store data locally and forward it when the path returns. Confirm the buffer did not overflow, and that the destination now holds the complete set with original timestamps.
  • Held locally on the instrument. PIC/S warns that temporary data storage on local systems before transfer to final storage creates an opportunity for data to be deleted or manipulated.11 Locally held data from a cutover window needs to be reconciled and its transfer verified.
  • Lost. If a flow drops data when blocked, a gap has to be assessed as a deviation, with impact on any batch in progress.

Knowing this before the window is what makes the rollback trigger meaningful. A flow that buffers for a day can tolerate a longer decision point than one that drops data at once.

Isolation Plans Use the Same Knowledge

Rollback plans cover a change that went wrong. Isolation plans cover an attack that is under way. The secure connectivity principles published in January 2026 by the UK NCSC with CISA, the FBI, and other partner agencies include, as their eighth principle, establishing an isolation plan.4 The guidance says the plan should be linked to wider business continuity plans and tested regularly. It also says that when planning large-scale isolation, it is important to identify critical data flows that must remain operational, and that some may need technical controls such as data diodes so they can keep running safely during a compromise.4

CISA’s 2025 fact sheet on primary OT mitigations makes a related point: “Practice and maintain the ability to operate OT systems manually.”13 For a GMP plant, isolating a zone raises the same quality questions as a failed cutover. Which records stop flowing? Can operators continue on paper batch records or local data capture? How will data captured during isolation be reconciled afterward? Annex 11 section 16 already expects provisions for continuity of critical processes in the event of a system breakdown, and says, “These arrangements should be adequately documented and tested.”8

The data flow register and the zone design give the isolation plan its content. Once a site has segmented, it can isolate one zone without stopping the whole plant, which is a large part of the reason to segment in the first place. The NCSC guidance also says isolation plans should take account of supplier contracts, for example whether a vendor can switch from remote support to attending in person.4 That is a quality agreement question as much as a security one.

Keeping Audit Trails and Time Sync Intact

Of all the things segmentation can break, time synchronization is the easiest to miss and among the most damaging to data integrity. When NTP (Network Time Protocol) traffic stops reaching a system, nothing fails at once. The system’s clock drifts a little each day. Weeks later, an investigator notices that an operator’s electronic signature appears to come before the step it approved, or that historian events and batch record entries for the same deviation are minutes apart and in the wrong order.

What the Regulations Expect

21 CFR Part 11 requires, for closed systems, the “Use of secure, computer-generated, time-stamped audit trails to independently record the date and time of operator entries and actions that create, modify, or delete electronic records.”12 Annex 11 section 12.4 says management systems for data and documents should record the identity of operators entering, changing, confirming, or deleting data, including date and time, and section 14 says electronic signatures should include the time and date they were applied.8

The PIC/S data integrity guidance is the most direct on clocks. In its section on system security for computerized systems, it says, “The operating system clock should be synchronized with the clock of connected systems and access to all clocks restricted to authorised personnel.”11 The same section expects firewall rules to be based on least privilege and periodic reviews of network security measures. A segmentation project that tightens firewalls but breaks clock synchronization meets one expectation in that list by failing another.

A Time Design for a Segmented Plant

NIST SP 800-82 treats time synchronization as its own topic. It says a common time should be used across all OT devices, that multiple time sources can reduce clock error and provide backup if the primary source is lost, and that authenticated NTP or secure Precision Time Protocol (PTP) can be used if there is a risk of malicious changes to network time.2 It also says, “Non-authenticated NTP is susceptible to spoofing and should be located behind the firewall,” and that time sources in the OT environment should be included in monitoring, with their logs forwarded to a log collection system.2

RFC 8633, the Internet engineering community’s best current practice for NTP, adds a point about numbers: “Operators who are concerned with maintaining accurate time SHOULD use at least four independent, diverse sources of time.” It also advises operators to monitor all time sources in use.14

4+Independent, diverse time sources recommended for accurate time in RFC 8633
8Secure connectivity principles for OT, the last being an isolation plan (NCSC and partners, 2026)
5Requirement areas in IEC 62443-3-2, from defining the system to documenting security requirements (IEC)

For a segmented GMP plant, these points lead to a design along these lines. Each site has internal time servers, ideally with their own reference such as a GPS-disciplined clock, so the plant keeps good time even when isolated from the enterprise. Time flows down the levels through defined conduits, with each zone taking time from the level above or from servers inside the plant, never directly from the internet. NTP is explicitly listed as a Tier 1 flow in the data flow register for every system that stamps GMP records. And the time servers themselves are monitored, so a failed or drifting source raises an alarm instead of being discovered in an investigation.

Controllers and HMIs deserve special attention. Many older devices cannot use NTP at all and keep their own clocks, set by hand or by a supervisory system. The register should record how each one gets its time, and the impact assessment should check whether that method survives the move.

Check Time at Every Cutover

Make clock checks a standard step in every wave’s test protocol:

  1. Before the window, record the clock offset of each system in the zone against the site reference.
  2. After the cutover, confirm each system is still synchronizing, and record the offset again.
  3. Generate a test event that is recorded in two systems, such as a batch record entry and the matching historian event, and confirm the timestamps agree within the site’s defined tolerance.
  4. Repeat the offset check at the end of the watch period, because drift appears over days, not minutes.

The tolerance itself should be defined by the site based on the processes involved. What matters is that it is written down, that it applies before and after the change, and that the test records show it was met.

Audit Trail Delivery

Many systems write their audit trails locally. Others write them to a central database, sometimes on a different server or in a different zone. For the second kind, the audit trail path is a Tier 1 flow and needs a rule, a test, and a “behavior when blocked” entry like any other. Some systems queue audit entries until the database is reachable. Others refuse further user actions until it is, which protects the record but can stop production. A few may lose entries. The project needs to know which is which before the firewall goes live.

Firewall and network device logs are a related but separate topic. NIST recommends that network and computing devices log events to support monitoring and incident response, with centralized log management to support retention and analysis.2 These logs are generally security records rather than GMP records, but they become very useful evidence during segmentation: they show which flows the new rules allowed and blocked, and when. Keep them for at least the length of each wave’s watch period, and longer if the site’s security policy requires.

Holding the Segmented State After Go-Live

The project ends. The network does not stop changing. New equipment arrives, vendors request access, and firewall rules accumulate. Without a plan for operation, a carefully segmented network drifts back toward a flat one, one approved exception at a time.

Treat the Rule Set as a Configuration Baseline

After the last wave, the approved rule set for each firewall becomes a controlled configuration item, the same as a validated system’s configuration specification. NIST SP 800-41 recommends commenting each rule so others can determine why it was made, and copying rule changes and their comments to the configuration management log.10 For a GMP site, the most useful comment on a rule is the data flow register ID it serves. That one link answers the two questions every later reviewer asks: why does this rule exist, and what breaks if we remove it?

Keep the Register Alive

The data flow register has to stay current to stay useful. Practical ways to do that:

  • Make “update the data flow register” a required step in change control for any new system, interface, or network change in an OT zone.
  • Add security requirements and a list of required data flows to equipment user requirement specifications, so the vendor states its communication needs before the equipment arrives. FDA’s white paper says vendors and manufacturer OT specialists should understand the extent and routing of system traffic within zones, between zones, and externally before deployment begins.1
  • Compare the register against what the network monitoring tools observe, on a schedule. A flow that appears in monitoring but not in the register is either undocumented or unauthorized.

Fold Rule Review Into Periodic Review

Annex 11 section 11 says periodic evaluations of computerized systems should include, where appropriate, security among other items.8 PIC/S expects regulated users to conduct periodic reviews of the continued appropriateness and effectiveness of network security measures.11 Both fit naturally with a periodic review of each zone’s rule set against the register: rules with no matching flow are candidates for removal, flows with no matching rule point to an error somewhere, and temporary rules past their end date are deviations.

A simple measure of drift: at each periodic review, count the rules in each zone’s firewall that do not trace to a current data flow register entry. The number should be zero, or each exception should have an open, time-limited change record. A count that rises from one review to the next is an early sign the segmented state is wearing away.

Who Owns What

Segmented networks fail most often at the handoffs between teams. A clear split of responsibilities helps:

RoleOwns
OT or network engineeringFirewalls, switches, rule sets, time servers, and the technical accuracy of the register
System ownersThe flows their validated system needs, the evidence that proves them, and approval of any rule change that affects them
QualityChange control, impact assessment approval, deviations from temporary rules, and periodic review
Information securityThe zone and conduit risk assessment, target security levels, monitoring, and the isolation plan
ValidationTest protocols for each wave and updates to validated documents

None of these roles is new. What segmentation adds is a shared document, the register, that all five use. That shared reference is what lets a security change move through a quality system without a fresh debate each time.

Conclusion

Segmenting a GMP plant network is a security project that changes qualified infrastructure underneath validated systems. Treated as a single large network change, it stalls in change control or, worse, gets pushed through and breaks records nobody knew depended on the old paths. Treated as a series of small, evidence-based cutovers, it can move steadily. The difference lies in the preparation: a data flow register that names every GMP flow and how it behaves when blocked, zones drawn around validation boundaries, one risk assessment that answers both security and quality questions, wave-level impact assessments, record-level tests in both directions, rollback triggers decided before the window, and time synchronization and audit trail delivery tested at every step. Done this way, segmentation also leaves the site with something it rarely had before: a current, reviewable account of what its validated systems need from the network.

Sakara Digital works with pharma and biotech organizations on the quality and validation side of OT modernization, including segmentation programs that need to satisfy security teams and inspectors at the same time. If you are planning zones and conduits for a GMP site and want an independent view of how to sequence the work with your quality system, we are happy to have that conversation.

For Further Reading