What AI forces us to confront in OT and critical infrastructure

COMMENTARY: The National Institute of Standards and Technology (NIST) in collaboration with the MITRE Corporation announced late last month that it's investing $20 million to establish two AI Economic Security Centers.

One will focus on advancing AI solutions for U.S. manufacturing. The other will work to secure critical infrastructure from cyber threats. The centers are part of a broader effort that includes NIST's planned AI for Resilient Manufacturing Institute, which will receive up to $70 million over five years.

[SC Media Perspectives columns are written by a trusted community of SC Media cybersecurity subject matter experts. Read more Perspectives here.]

It’s an important moment. Not because of the dollar figure, but because it forces us to confront realities that have long been deferred in favor of operational continuity. It’s also important because of what it represents. We’re now actively introducing AI into environments where a wrong decision can mean physical consequences, operational disruption, or cascading failures across critical systems.

Moving forward, AI systems will not just observe, but recommend, decide, and act at a speed and scale humans can’t match. In industrial environments, that's both exciting and uncomfortable. Productivity gains are real. So are the risks. And those risks demand we take a hard look at whether we're actually ready for this shift.

This pressure has emerged now for a reason. Industrial environments are more connected than they have ever been, driven by remote operations, distributed workforces, and the need to operate assets across wider geographic footprints. At the same time, AI capabilities have matured from niche analytics into tools that organizations increasingly expect to operate in real time. Add workforce constraints and growing operational complexity, and we have convergence that makes AI adoption feel less optional and more inevitable.

NIST’s announcement reflects that reality: It's no longer a discussion on whether AI will get introduced into these environments. The conversation has shifted to whether we will do it deliberately or by necessity, and whether existing operating and security models can handle it.

Most OT environments were built on the following assumptions that no longer hold:

Those assumptions were fragile even when only humans were involved. When we introduce AI systems that can act autonomously across these environments at machine speed, those cracks become fault lines.

If AI systems are allowed to interact with industrial control environments, we have to be far more explicit about who or what is allowed to act, under what authority, for how long, and with what level of visibility and accountability.

AI does not remove human responsibility from these systems. In many ways, it heightens it. When actions are automated or recommended at machine speed, accountability can become harder to assign unless it's designed into the system from the start. In OT environments, responsibility cannot be abstracted away behind algorithms or tools. Someone must still own outcomes, understand why actions were taken, and be able to intervene when behavior diverges from intent. Without that clarity, automation risks creating ambiguity at exactly the moments when decisiveness matters most.

This isn't an argument against AI. It's an argument for readiness, especially as industrial environments become more remote and more connected.

Remote access makes this more urgent. OT environments increasingly rely on remote connectivity for monitoring, maintenance, and operations. When AI agents gain remote access, they don't operate on human timescales or with human limitations. A single misconfiguration or compromised identity can now execute hundreds of actions across dozens of systems before anyone realizes what's happening.

The real opportunity in NIST's AI Economic Security Centers isn't just better models or smarter agents. It's the chance to rethink the foundations: identity, access, containment, and traceability. These are not new ideas in security, but AI makes them unavoidable.

By pushing that rethinking forward, these centers create space for experimentation, learning, and iteration, recognizing that AI adoption in these environments will require careful testing as much as confidence.

Zero-trust in OT isn't about adopting a framework: it's about recognizing that when we give systems the ability to act autonomously in physical environments, control becomes non-negotiable. Intelligence without control is not progress.

AI will reshape how industrial systems are designed, operated, and defended. So will that happen on our terms or the AI agents? And, can we adapt systems built for stability to meet threats moving at machine speed?

That's the conversation this announcement opens. And it's one our industry needs to have now.

Bill Moore, founder and CEO, Xona

SC Media Perspectives columns are written by a trusted community of SC Media cybersecurity subject matter experts. Each contribution has a goal of bringing a unique voice to important cybersecurity topics. Content strives to be of the highest quality, objective and non-commercial.