The integration of Artificial Intelligence (AI) into manufacturing environments is no longer a futuristic aspiration; it is an active, ongoing transformation. However, as Industrial AI moves from experimental pilot programs to the factory floor, it brings with it a complex web of security, safety, and architectural challenges.
During the "It Can’t Happen to Me – Until It Does" session at the Industrial AI Summit 2026, industry experts Gary Tillery of Skkynet and Matt Morris of EverLine addressed the critical disconnect between the promise of AI and the harsh realities of Operational Technology (OT) security. The panel emphasized a sobering truth: in the current industrial landscape, there is no single "golden standard" for AI, and organizations that rely on traditional IT-centric security models are inherently vulnerable.
The Architectural Imperative: Why IT/OT Convergence is a Double-Edged Sword
The core of the security challenge lies in the fundamental difference between Information Technology (IT) and Operational Technology. While IT prioritizes data confidentiality and uptime, OT prioritizes human safety, process integrity, and physical reliability.
Gary Tillery, CEO of Skkynet, argues that the nuance of OT risk is not a communication hurdle—it is a consequence-driven reality. "A bad change in an OT environment can stop a production line or, worse, create a catastrophic safety event," Tillery explained. "The installed base of industrial machinery is often old, proprietary, and vendor-specific. In this environment, the ‘safe’ action is frequently to leave a vulnerable asset running until a planned outage, simply because the risk of patching outweighs the risk of the vulnerability."
When organizations attempt to layer AI onto these legacy systems, they often fail to account for the fragility of the underlying architecture. The panel stressed that "awareness training" is insufficient; boards and plant managers require a structured, defensible, and actionable stack that segregates AI-driven advisory tools from the hard-wired logic of safety systems.
Decoding the Standards Stack: A Multilayered Approach
One of the most frequent questions from the Industrial AI Summit audience centered on the ambiguity of existing standards. Attendees asked: "Which standards actually govern AI in manufacturing?"
Tillery clarified that the industry is currently operating in a "stack" rather than a singular framework. Because agentic and generative AI models can act as planners that select their own tools, they present a novel risk profile that older, static standards did not anticipate. To secure these systems, organizations must integrate four distinct layers:
1. The Industrial Interoperability Layer
IEC 62541 (OPC UA) serves as the practical bedrock. By utilizing standardized information models, AI can "read" machine states rather than raw, context-free tags. Furthermore, IEC 63278 (Asset Administration Shell) and the RAMI 4.0 stack provide a standardized digital twin representation for each asset, ensuring that AI retrieval is grounded in validated asset data.
2. The Control and Security Layer
This is where the IEC 62443 series remains paramount. Regardless of how sophisticated an AI model becomes, it must be subject to the same zone-and-conduit principles as any other OT interface. NIST SP 800-82 Rev. 3 provides the necessary roadmap for applying these controls to the OT environment. Crucially, the AI model must never hold standing credentials to Programmable Logic Controllers (PLCs) or data historians.
3. The Functional Safety Layer
The panel was emphatic: AI must never be allowed to cross the boundary of functional safety. Safety functions must remain governed by established standards such as IEC 61508, 61511, or 62061 / ISO 13849. If a model is used for optimization, it must operate as an advisor, with all actual commands processed through a broker that enforces human-gated, task-scoped authorization.
4. The Trustworthiness and Lifecycle Layer
ISO/IEC 42001:2023 serves as the primary standard for a certifiable AI management system. When paired with the NIST AI Risk Management Framework (AI RMF), these provide the governance structure required to ensure that an AI system remains as secure and predictable as the day it was commissioned.
The Phishing Fallacy: Rethinking the IT/OT Boundary
A central tension in modern manufacturing is the belief that firewalls are enough to stop a digital intruder. Matt Morris, VP of Security and Resilience at EverLine, challenged this premise directly: "If a single phishing click can reach your OT environment, you don’t have a phishing problem; you have an architectural problem."
The "Assume Compromise" Mindset
Morris argued that security design must proceed under the assumption that the IT environment is already compromised. If the plant floor is architected such that an IT breach allows for lateral movement into the OT network, the organization is effectively one email away from an operational shutdown.
This is compounded by "operational dependency." Referring to the Colonial Pipeline incident, Morris pointed out that the OT systems were not directly breached; rather, the IT systems were locked, and the company was unable to manage the billing and tracking of product. Consequently, they had to shut down the pipeline as a safety precaution. If OT operations cannot function in the absence of IT connectivity, the boundary between the two is, in effect, a fiction.
Proving Control Effectiveness
To move beyond theoretical security, companies must adopt rigorous, proactive verification:
- Segmented Testing: Conduct regular penetration testing that specifically targets the "conduits" between zones.
- Operational Resilience Drills: Test the ability of the plant to "go dark"—operating in an air-gapped or disconnected state—to ensure that IT failure does not necessitate a total production halt.
- Credential Auditing: Ensure that AI agents or third-party tools have short-lived, task-scoped credentials that leave an immutable audit trail.
Toward a New Network Architecture: Beyond the Internet
As the limitations of the public internet become more apparent in the context of high-stakes industrial control, the question arises: should manufacturers build dedicated, "walled-off" networks similar to military-grade systems?
Gary Tillery noted that while replacing the entire network is rarely feasible for legacy facilities, specific software solutions can simulate the benefits of a secure, dedicated network. Tools like Skkynet’s Cogent DataHub allow for outbound-only data flows. By eliminating open inbound ports in the firewall, organizations can remove the attack surface that typically allows external actors to probe the OT network.
"The goal is to design the network to maximize the security of one domain from infecting another," Tillery explained. "By combining software-based outbound-only gateways with hardware-based solutions like data diodes, you create a tiered architecture where the AI can observe and advise without ever being able to initiate a command-and-control connection back into the critical safety loop."
Implications for the Future of Industrial AI
The 2026 Industrial AI Summit made it clear: the era of "plug-and-play" AI in industrial settings is over. The future belongs to organizations that treat AI integration as a disciplined engineering challenge rather than a software deployment.
Key Takeaways for Stakeholders:
- Governance Over Convenience: Adhere to ISO/IEC 42001 and NIST AI RMF to ensure that AI systems are not only efficient but also auditable and compliant with evolving regulations, such as the EU AI Act.
- Safety is Non-Negotiable: Never allow AI, regardless of its sophistication, to override the hard-coded functional safety standards (IEC 61508/61511).
- Architecture is the First Line of Defense: Stop relying on firewalls to act as the sole barrier. Implement micro-segmentation, strictly enforced least-privilege access, and outbound-only communication patterns.
- Plan for the Worst: Assume IT will be compromised. Build industrial processes that can survive a temporary or permanent loss of connectivity to enterprise IT systems.
As the industry continues to evolve, the distinction between "smart" and "secure" will blur. Organizations that prioritize the structural integrity of their OT environments while layering on the analytical power of AI will find themselves at a distinct competitive advantage—not just in terms of efficiency, but in the resilience of their most critical assets.
The panel, featuring insights from Tillery, Morris, and contributions from industry peers at Black & Veatch and GrayMatter, serves as a vital call to action: the time to redesign the industrial architecture for the AI age is now, before the next "it can’t happen to me" moment becomes an operational reality.
