Table of Contents
Adding AI to a plant does not replace your OT security programme; it adds new assets, data flows, vendors and failure modes to it. Treat every model, data pipeline and vendor connection as a new OT component: assess its risk with ISA/IEC 62443, keep it out of the control path, monitor it, and make sure operators can run without it.
Key takeaways
- Joint guidance from CISA and eight partner agencies lists AI risks for OT, from manipulated data to model drift and overreliance, and says LLMs "almost certainly should not be used to make safety decisions" (CISA et al., 2025).
- Collecting OT data in one place for AI creates a target: the same guidance warns that attackers can use it for a more targeted engineering impact.
- Most new exposure comes from paths: data pipelines, vendor remote access and AI features that need the internet. Approve each one and route it through the existing DMZ or jump host.
- Add AI failure states to incident response and keep the ability to operate manually, one of CISA's five primary OT mitigations (CISA, 2025).
- In Algeria, a national information systems security framework has existed since Presidential Decree 20-05 of 2020. Confirm with counsel which texts apply to your site.
What does AI change in OT security?
The old risks stay. The CISA-led guidance, written with US, Australian, Canadian, German, Dutch, New Zealand and UK agencies, states that "traditional cybersecurity risks remain within AI systems". AI adds its own failure modes on top (CISA et al., 2025).
Attacks on industrial systems in the region are not hypothetical. NIST SP 800-82r3 lists Shamoon, which made systems at Saudi Aramco unusable in 2012, and TRITON, malware that targeted the safety instrumented system of a Saudi petrochemical facility in 2017 (NIST SP 800-82r3).
The table below follows the risks in Table 2 of the CISA-led guidance. The plant examples and first controls are ours.
| Risk named by CISA | What it can look like in a plant | First control |
|---|---|---|
| Data, models or software manipulated to give wrong results or bypass safeguards | Altered training data teaches an anomaly model that an abnormal vibration is normal | Restrict who can change training data and models; test before every release |
| Data quality, and OT data gathered in one place | A data lake holding historian extracts and P&IDs becomes a map of the plant | Classify the data; encrypt it and log every access |
| Model drift | Predictions degrade after a process change or new feed | Monitor outputs against known-good bounds; define when to fall back |
| False alarms and operator cognitive load | A detection model floods the panel with alerts | Set alarm targets in the business case; tune before production |
| Unreliable output and hallucination | An assistant quotes a wrong trip value from a manual | No safety decisions by an LLM; cite sources; human approval |
| Overreliance and skill erosion | Operators stop cross-checking local gauges | Training, procedures and manual operation drills |
Protect engineering configuration data first: the guidance says network diagrams, asset inventories, logic diagrams and safety information have "enduring value" to attackers.
Where do new attack paths come from?
Three paths deserve most of the attention.
Data pipelines. NIST describes Havex, a remote access Trojan that used the OPC standard to gather information about connected ICS devices (NIST SP 800-82r3). AI data pipelines often read the same protocols, so the collector must be read-only and push data out of OT rather than give the AI system inbound access. Our article on the Purdue model and where AI fits sets out the data-flow rules, and data diodes and one-way transfer covers hardware-enforced options.
Vendor remote access. AI systems need updates and support. NIST says remote access to OT should be justified, limited to the business need, never allowed to circumvent safety or security controls, and that multi-factor authentication should be considered. The CISA-led guidance adds that new connection paths, such as remote access, must be approved and pass through the existing DMZ or jump host.
AI features that need the internet. CISA notes that some OT devices now ship with built-in AI that may require internet connectivity, a new outbound path from OT. The first of CISA's five primary mitigations for OT is to remove OT connections to the public internet (CISA, 2025).
Self-hosted AI servers need the same discipline. Cisco found 1,139 Ollama servers exposed on the internet, 214 of them serving models (Cisco, 2025). An unauthenticated inference server on a plant network is the same mistake on a smaller scale; see securing an on-premise LLM.
How do you fit AI into an existing OT security programme?
The CISA-led guidance asks owners to embed AI in their existing risk evaluation, mitigation and monitoring processes, not to build a parallel track.
- Check the business case first. CISA asks whether an established capability already meets the need. Write security requirements next to the performance goal; its fictional generator example includes keeping data collection under 80% of network bandwidth.
- Add AI to the asset inventory. NIST stresses that an accurate inventory supports risk assessment and vulnerability management. CISA asks owners to inventory AI components and the components that depend on them.
- Define zones and conduits. ISA-62443-3-2 requires defining the system under consideration, partitioning it into zones and conduits, assessing risk for each, setting a target security level and documenting the requirements (ISA-62443-3-2 preview). Give the AI system its own zone and treat each data flow as a conduit with an owner.
- Add AI techniques to the threat model. CISA recommends using MITRE ATLAS alongside ATT&CK. ATLAS is "a public knowledge base of adversary TTPs targeting AI systems" (MITRE ATLAS). The voluntary NIST AI Risk Management Framework can structure the governance side.
- Test away from production. Start on test infrastructure, move to realistic non-production systems such as hardware in the loop, and avoid exposing production data in test environments.
- Keep the AI out of the control path. Limit active control by AI without a human in the loop; send recommendations through change management. See human-in-the-loop for AI agents.
- Set the fallback threshold. CISA asks owners to define when the plant defaults back to non-AI systems, for example when outputs fall below performance and safety thresholds.
What should you ask AI and OT vendors?
Combine the AI questions from the CISA-led guidance with the twelve security elements of the Secure by Demand guide for OT buyers (CISA et al., Secure by Demand, 2025).
| Question | Why it matters |
|---|---|
| Which AI features does the product contain, and what do they do? | You cannot assess what is not declared |
| Will you provide an SBOM and say where the models are hosted? | Shows the software supply chain and any outside dependency |
| Will you notify us if the AI is found to give improper advice? | Goes beyond normal vulnerability disclosure |
| Will you train models on our operational data? Where is it stored, and over which paths? | Protects IP and keeps data under your control |
| Can it run on premises without continuous internet access? | Avoids a new outbound path from OT |
| Can our operators disable AI features, and under which conditions? | Keeps control with the operator |
| Can we operate and maintain it without depending on you, including through local engineering firms? | Secure by Demand's ownership test |
Our AI project RFP guide shows how to turn these questions into tender requirements.
How do you monitor an AI system in OT?
The CISA-led guidance recommends:
- Log inputs and outputs of AI components, with an AI identity distinct from other machine and user identities, so decisions can be traced.
- Collect flow and access logs for AI endpoints and track data egress by asset and identity.
- Watch the outputs statistically. Its predictive maintenance example applies statistical process control to flag outliers in predicted downtime.
- Show AI health where operators look, ideally in existing HMI views.
- Red team the AI system regularly, and review its KPIs with vendors, operators and the governance board.
What changes in incident response?
AI creates new failure states. The guidance asks owners to add them, including how to bypass or replace the AI system, to functional safety and incident response procedures, covering both attacks on the AI and plain AI failure.
People are part of the fallback. The guidance suggests training teams to check AI output against alternative sensors, including their own senses, and to know what to do when it is invalid. CISA's primary OT mitigations include "practice and maintain the ability to operate OT systems manually". Switch off a model operators rely on in a drill, before it fails on its own.
What does Algerian regulation say?
Algeria's National Strategy of Information Systems Security states that Presidential Decree 20-05 of 20 January 2020 created a national framework made up of a National Council and the Information Systems Security Agency (ASSI). The strategy names energy and water resources among the sectors whose compromise would have a major impact, and puts protecting critical infrastructures in its first axis (National Strategy, ITU repository).
According to press reports, Presidential Decree 26-07 of 7 January 2026, published in the Official Journal on 21 January, requires public institutions, administrations and bodies to set up a cybersecurity structure separate from the one that runs their information systems. Its tasks include drafting the security policy, mapping digital security risks, continuous monitoring and remediation plans; incidents must be reported to the competent authorities without delay, and the decree calls for security and confidentiality clauses in subcontracting contracts (Algérie360, 2026). If the decree applies to your organisation, an AI vendor contract may fall under that requirement.
Read the texts in the Official Journal and confirm with legal counsel how they apply to your organisation, especially for public economic enterprises.
What we have seen at KABAS AI
DALILE runs completely offline on the client's own servers, with encryption, access controls and audit logging, and connects to systems such as SAP, LIMS and SQL databases. Placing such a system near a plant raises the questions above, so we recommend three steps.
Bring in the OT security owner at the start. Each data source the AI system will read should have an approved conduit before the pilot, not after it.
Start with read-only, advisory use cases. A document assistant or maintenance recommendations from historian data add little risk to the control system. Anything that writes to it comes later, if at all, under its own change process.
Write the fallback before go-live. Decide who switches the AI off, how operators work without it and which threshold triggers that decision.
Our information systems team integrates AI with existing systems, and our consulting team runs readiness assessments and sets up AI governance. Related: integrating AI with SCADA and OT systems and, for fully isolated sites, air-gapped AI deployment.
Sources
- CISA, ASD's ACSC and partners, Principles for the Secure Integration of Artificial Intelligence in Operational Technology, 3 December 2025
- CISA, Primary Mitigations to Reduce Cyber Threats to Operational Technology, May 2025
- CISA and partners, Secure by Demand: Priority Considerations for Operational Technology Owners and Operators when Selecting Digital Products, January 2025
- NIST SP 800-82r3, Guide to Operational Technology (OT) Security, September 2023
- NIST, AI Risk Management Framework
- ISA, ANSI/ISA-62443-3-2-2020 preview
- MITRE ATLAS data repository
- Cisco, Detecting Exposed LLM Servers: A Shodan Case Study on Ollama, September 2025
- Algeria, National Strategy of Information Systems Security (ITU National Cybersecurity Strategies Repository)
- Algérie360, Cybersécurité : l'Algérie se dote d'un cadre réglementaire pour faire face aux cybermenaces, January 2026
Frequently asked questions
Does adding AI to a plant increase its cyber risk?
It adds to the attack surface. The joint guidance led by CISA notes that AI adds complexity and can create new attack paths, especially when it brings remotely accessible or internet-exposed connections. The added risk can be limited if the AI system is treated as a new OT component: placed in its own zone, given read-only data, monitored and covered by incident response.
Should an AI model be allowed to write setpoints to a PLC or DCS?
Not as a starting point. The CISA-led guidance recommends limiting active control of OT by AI without a human in the loop, and says LLMs almost certainly should not be used to make safety decisions. The usual pattern is a recommendation that an operator reviews and applies through the existing change management process.
Is ISA/IEC 62443 enough to secure an AI system in OT?
It gives the method: define the system, split it into zones and conduits, assess the risk of each and set a target security level. It does not list AI-specific attacks. CISA and its partners note that current AI standards are written mainly for IT, and suggest adding AI techniques to threat models, for example with MITRE ATLAS.
What should an AI or OT vendor contract say about plant data?
It should say whether the vendor may train models on your operational data, where data is stored, which communication paths it uses and how it is encrypted. The CISA-led guidance also suggests requiring a software bill of materials, disclosure of where models are hosted, notification if the AI is found to give improper advice, and operator control over turning AI features off.
Does on-premise AI remove the need for vendor remote access?
Not always, because support and updates still have to reach the system. NIST SP 800-82r3 says remote access to OT should be justified, limited to the business need, never allowed to bypass safety or security controls, and that multi-factor authentication should be considered. Updates can also be carried in through a controlled offline process.
Which Algerian cybersecurity rules apply to AI projects in industry?
Presidential Decree 20-05 of 2020 created the national information systems security framework and the Information Systems Security Agency. According to press reports, Presidential Decree 26-07 of January 2026 requires public institutions, administrations and bodies to set up dedicated cybersecurity structures and add security clauses to subcontracting contracts. Whether and how these texts apply to your company is a question for legal counsel.



