Key takeaways
- Design AI data flows so OT pushes data out. Avoid any persistent inbound path from the cloud or IT into Level 2 or below.
- An edge inference box is an OT asset: inventory it, zone it, harden it and put it under management of change.
- Model updates are software changes. Decide who can push them, how they are tested and how you roll back.
- Data poisoning and model drift are real but mostly quality and availability risks. Plan for the AI being wrong, not only for the AI being hacked.
- Every AI-assisted process needs a defined fallback to traditional automation or manual operation.
AI is arriving in plants through many doors at once. A vision system on a packing line. A predictive maintenance box bolted into a panel by the OEM. A cloud analytics platform that wants historian data. A copilot that reads alarm logs. Each one is a new piece of software and a new network path into or out of your operational technology. This guide covers how to secure them without stopping the projects.
What changes when AI enters the OT network
From a security point of view, most industrial AI is not exotic. It is software running on a server, an industrial PC or a cloud service, connected to data sources you already have. The new risks come from three places:
- New connectivity. AI projects often need data from Level 1 and 2 to reach a cloud platform, and sometimes need results to flow back down. That is a new conduit.
- New software supply chain. Models, Python runtimes, container images and vendor update channels. Many of these are updated far more often than your PLC firmware.
- New failure modes. A model can be wrong, drift as the process changes, or be fed bad data. If it influences setpoints or quality decisions, that matters for product, safety and uptime.
In December 2025, CISA, Australia's ASD's ACSC and partner agencies in the US, Canada, Germany, the Netherlands, New Zealand and the UK published joint guidance, Principles for the Secure Integration of Artificial Intelligence in Operational Technology. It sets out four principles: understand AI, consider AI use in the OT domain, establish AI governance and assurance, and embed safety and security practices. It is short and worth reading. The rest of this article turns those principles into plant-floor decisions.
Does this apply to you?
Yes, if any of these are true:
- An OEM or integrator has installed an "AI", "smart" or "predictive" module on a machine you own.
- You stream historian, MES or sensor data to a cloud analytics or AI service.
- A vision inspection system uses a trained model to pass or reject parts.
- Anyone is piloting a generative AI assistant that reads OT data (alarms, work orders, manuals).
If you have no idea whether any of these apply, start with an AI use-case inventory and your OT asset inventory. Shadow AI on the shop floor usually arrives inside vendor equipment.
Architecture: where AI workloads should sit
The joint guidance recommends push-based architectures: data leaves the OT network so AI can use it, without persistent inbound access into OT. That one rule eliminates most of the bad designs we see.
Cloud AI / analytics platform
^ (HTTPS out, initiated from DMZ)
|
+---------+---------------------------+
| L3.5 Industrial DMZ |
| historian replica | MQTT broker |
+---------+---------------------------+
^ (one-way or outbound only)
|
+---------+---------------------------+
| L3 Site ops: historian, MES |
+---------+---------------------------+
^
+---------+---------------------------+
| L2 Supervisory: SCADA, HMI, |
| edge inference box (own zone) |
+---------+---------------------------+
^
| L1/L0 PLCs, drives, sensors |
Practical placement rules:
- Edge inference boxes get their own zone, or at least their own VLAN and firewall rules, at Level 2 or 3. They should read from a defined source (OPC UA server, historian, camera) rather than browsing the control network.
- Cloud connections terminate in the DMZ, not at Level 2. The cloud never initiates a session into OT.
- Writes back to control are the exception. If AI output must change a setpoint, route it through the HMI or SCADA with operator confirmation, or through a narrowly scoped, logged interface with hard limits enforced in the PLC.
- Training happens outside OT. Models are trained in IT or cloud, then delivered as a release package. No training jobs pulling live data from Level 1.
For the reasoning behind these levels, see the Purdue model and AI. For the formal method, see zones and conduits.
Data diodes vs one-way gateways vs brokered MQTT
These three get confused. They are not interchangeable.
| Option | How it works | Strength | Trade-offs | Good fit |
|---|---|---|---|---|
| Hardware data diode | Physical one-way link, often optical. Nothing can return. | Direction enforced in hardware; misconfiguration cannot reverse it | Cost; no acknowledgements; protocols need replication software on both sides | High-consequence processes, safety-adjacent data, telemetry to monitoring |
| One-way gateway (diode plus replication software) | Diode hardware with software that replicates historians, OPC servers or files to the outside | Same guarantee, plus usable data on the IT side | Vendor lock-in; replication configuration must be maintained | Feeding a cloud AI platform from a critical site |
| Brokered MQTT / outbound-only | Edge publishes to a broker in the DMZ; consumers subscribe from IT or cloud | Cheap, flexible, fits unified namespace designs | Relies on firewall rules and broker hardening; TCP is bidirectional | Most mid-market analytics and AI pilots |
The 2026 Secure connectivity principles for operational technology from the UK NCSC, ASD's ACSC and partners make a useful point: a diode enforces direction in hardware, but a diode alone is not a full cross-domain solution. Running diodes in both directions to fake a two-way link is an anti-pattern.
Plant-floor reality: Most mid-market plants will run brokered, outbound-only flows through a DMZ for years. That is acceptable if the broker is hardened, authenticated, patched and monitored, and the firewall only allows OT-initiated connections. Reserve diodes for the lines where a breach means a safety or major production event.
Model integrity and update control
A model file is code that makes decisions. Treat it that way.
Who can push model updates. Write it down. A typical answer:
- The vendor or data science team builds and tests the model outside OT.
- They deliver a signed, versioned package with release notes and validation results.
- The OT engineer or process owner approves deployment through management of change (MOC).
- Deployment happens in a planned window, with a recorded rollback version.
- Nobody, including the vendor, can push directly to production edge devices over the internet.
Integrity controls. Hash or sign model artifacts. Keep a model registry with version, training data snapshot, owner and approval date. Lock down write access on edge devices. Log model version changes the same way you log PLC program changes.
Vendor update channels. Many edge AI products phone home for updates by default. The joint guidance asks operators to confirm products can run on-premises or without constant internet access, and to keep control over enabling or disabling AI features. Put that in the contract. Our OT/AI vendor security questionnaire and the AI vendor due diligence guide cover the questions.
Adversarial attacks and data poisoning: a realistic view
NIST AI 100-2 E2025 describes the main attack classes against machine learning: evasion (crafted inputs at inference time), poisoning (corrupting training data or the model), privacy attacks, and, for generative AI, misuse and prompt injection. All are real research topics. In a mid-market plant, the risk ranking looks like this:
- Accidental bad data. A drifting sensor, a mislabeled reject bin, a recipe change nobody told the data team about. This is the most likely cause of a model degrading QC or maintenance decisions.
- Ordinary compromise of the AI host. Ransomware or a remote access compromise on the edge box or cloud tenant. This is a classic IT/OT problem, and the AI is just the entry point or the victim.
- Deliberate poisoning or evasion. Possible, especially from an insider or a compromised supplier pipeline, but it requires knowledge of your process and model. Lower likelihood, potentially high impact.
- Prompt injection against LLM assistants. Relevant if an assistant reads untrusted text (emails, supplier documents, work orders) and can take actions. Keep such assistants read-only in OT.
Watch out: Do not let anyone sell you exotic adversarial ML defenses before you have patching, MFA and segmentation on the AI host. The boring controls stop the likely attacks.
Controls that address both the accidental and deliberate cases: data lineage for training sets, validation against a held-out set before release, performance monitoring in production, and defined thresholds that switch back to non-AI operation.
Fallback and human oversight
The joint guidance is clear that AI-enabled processes should be able to revert to traditional automation or manual operation, and that AI failure states belong in functional safety and incident response plans. On the floor this means:
- AI must never sit in the safety function. The SIS stays independent.
- For each AI use case, document what happens when the model is offline or wrong. "Line runs on fixed recipe" or "vision reject reverts to manual inspection at station 4" are good answers.
- Operators can see when AI is advising or acting, and can override it.
- Define drift and confidence thresholds that trigger fallback automatically or alert the process owner.
Include an AI failure in your OT incident response tabletops.
Monitoring
You cannot defend what you cannot see. Minimum monitoring for AI in OT:
- Network: flows from edge AI devices and the DMZ broker baselined in your OT monitoring tool or firewall logs. Alert on new destinations or inbound attempts.
- Host: edge devices sending logs to your SIEM or log collector (via the DMZ). Track logins, software changes and model version changes.
- Model: accuracy, false reject rate, prediction confidence and input data ranges, reviewed by the process owner on a set cadence.
- Vendor access: every remote session to an AI device recorded and tied to a ticket. See secure remote access.
What good looks like
- Every AI system touching OT is in the asset inventory and the AI use-case register, with an owner.
- Each has a zone, a documented conduit and an architecture that pushes data out.
- Model updates follow MOC with signed packages and a rollback version.
- Fallback mode is documented, tested and known to operators.
- Monitoring covers network, host and model performance.
Effort and cost
For a single-site manufacturer with two or three AI use cases, expect roughly two to four weeks of part-time OT and IT engineering effort to document flows, add firewall rules, harden edge devices and write the update procedure. Firewall and broker hardening is mostly labor. A hardware one-way gateway is a capital purchase, so scope it to the lines that justify it.
Common mistakes
- Letting the OEM's edge box connect straight to its cloud over the plant's general internet breakout.
- Running the AI pilot on a laptop under someone's desk on the control VLAN.
- No owner for the model once the data scientist or vendor leaves.
- Treating a model retrain as "just data" and skipping MOC.
- Assuming the AI vendor's SOC 2 report covers how their device behaves inside your OT network.
Monday-morning checklist
- List every AI or "smart" system connected to OT, including OEM modules. Add them to the asset inventory.
- For each, draw the data path. Mark any inbound connection into Level 2 or below.
- Close or broker any inbound path. Move cloud connections to the DMZ.
- Write a one-page model update procedure and get the plant manager to sign it.
- Document the fallback mode for each AI use case and brief the shift leads.
- Send the vendor security questionnaire to each AI supplier.
This article is technical guidance, not legal advice. Confirm any regulatory obligations with counsel.
Sources and further reading
- Principles for the Secure Integration of Artificial Intelligence in Operational Technology (CISA, ASD's ACSC and partners, Dec 2025)
- Principles for the secure integration of AI in OT (ASD's ACSC)
- Secure connectivity principles for operational technology (ASD's ACSC, NCSC-UK and partners, Jan 2026)
- NIST AI 100-2 E2025, Adversarial Machine Learning: A Taxonomy and Terminology of Attacks and Mitigations
- NIST SP 800-82 Rev. 3, Guide to Operational Technology (OT) Security
This guide is general information for manufacturers, not legal advice. Confirm regulatory obligations with qualified counsel.
FAQ
Frequently asked questions
Is it safe to connect an AI analytics platform directly to my PLCs?
Usually not as a direct connection. Read data through an existing historian, OPC UA server or edge gateway in a supervisory or DMZ zone, and keep the AI platform out of the control zone. If the platform needs to write setpoints, route those writes through operator approval and your normal change process, and log every write. Direct cloud-to-PLC paths create a standing attack path into the process.
Do I need a data diode to send OT data to the cloud for AI?
Not always. A hardware data diode gives the strongest guarantee that nothing comes back in, and it suits high-consequence processes. Many mid-market plants use a hardened DMZ with outbound-only connections from OT, a broker or historian replica in the DMZ, and tight firewall rules. Pick based on consequence: safety and high-value continuous processes justify the diode.
What is data poisoning in a manufacturing context?
Data poisoning means corrupting the data a model learns from so it behaves badly later. In a plant, this could be tampered labels on a vision inspection dataset or manipulated sensor history used to train a predictive model. The more common real-world version is accidental: a miscalibrated sensor or mislabeled scrap that quietly degrades a model. Controls are the same: data lineage, change control and validation before release.
Who should be allowed to push AI model updates to the shop floor?
Only named people through a controlled path, with OT approval. Treat a model update like a PLC logic change: a change request, test evidence, a scheduled window, a rollback plan and sign-off from the process owner. Vendors should not be able to push updates automatically into production. Require them to deliver signed packages that your team stages and deploys.
Related guides
OT Security
IEC 62443 zones and conduits: a walkthrough for a mid-size plant
Zones group assets that share security requirements; conduits are the controlled paths between them.
Guide · 8 min read
OT Security
The Purdue model and AI: where analytics and cloud fit in your plant
The Purdue model is still the clearest way to explain where things sit in a plant network and which way data should flow.
Guide · 7 min read
AI Governance
AI vendor due diligence for industrial software and OEM equipment
Most AI in a plant arrives through vendors: MES, CMMS, QMS, vision integrators and OEM machines.
Guide · 7 min read
Data Foundations
Edge vs cloud AI in manufacturing: a governance and security decision guide
Run AI at the edge when it needs fast, deterministic responses, high-volume raw data, or must keep working without the internet.
Guide · 6 min read