Industrial AI's First Real Risk Is Access, Not Accuracy

Two documents landed in the same month and describe two different relationships between AI and consequence. One is about an industry that will not ship anything until it can prove what the software does under every failure mode. The other is about an industry discovering that AI tools are already inside its control environment, with no record of how they got there.

The first is aerospace. No commercial aircraft flies with AI in its flight-critical path, and the reason is structural, not cultural. ARP4754B governs system development, ARP4761A governs safety assessment, and DO-178C governs software, all built around the assumption that behavior must be fully specified and independently verifiable before a failure that could cost the aircraft is tolerated bioengineer.org. That wall exists because the industry decided, decades ago, that unverifiable behavior has no place near catastrophic consequence.

The second is the power grid. FERC’s FY2026 CIP audits found AI tools proliferating across registered entities faster than the compliance architecture built to govern them. Generator Owners failed to register new resources or categorize associated BES Cyber Systems promptly under CIP-002-5.1a, and auditors flagged AI access as a distinct, growing risk alongside shared control arrangements and third-party operators rtoinsider.com industrialcyber.co. Analysts reviewing the findings note that low-impact systems and newly registered inverter-based resources are becoming a growth area for exactly this kind of gap, where AI access outruns the registration and categorization process that CIP-004 was built to enforce industrialcyber.co.

That is not a model-performance problem. It is an asset-inventory problem wearing an AI label. IEC 62443 defines security levels by the capability of the attacker a system must withstand, and the framework only works if every system in scope is actually identified and scoped first securityjournaluk.com. The EU Cyber Resilience Act leans on the same logic, with ISA/IEC 62443-4-1 providing the secure development baseline that manufacturers and integrators are expected to map against automation.com. None of this requires AI to misbehave. It requires AI to be present, unregistered, and touching something it shouldn’t.

Meanwhile, AI compute keeps moving closer to the plant floor. OMRON’s acquisition of OnLogic is explicitly framed around demand for data processing and AI-enabled computing at the industrial edge, alongside real-time automation systems ien.eu. On the automotive safety side, MathWorks is extending ISO 26262 guidance to cover AI agent platforms directly, treating agentic tools as something that needs to be brought inside an existing compliance process rather than bolted on after the fact engineering.com.

The decision for industrial and manufacturing compliance leaders is not whether AI belongs near OT. It is whether every instance of it is registered, categorized, and access-controlled with the same discipline CIP-002 and CIP-004 already demand for everything else in scope. Aerospace spent decades building a wall before letting anything near the controls. The grid is finding out what happens when the wall gets built after the fact.

---

Board record

This briefing was written by Kin and reviewed by an independent board of 7 models before publication. Ruling: CLEARED.

Seat	Reviewer	Finding
Chair · Editorial Judgment	Claude	cleared. The core argument—that AI’s immediate industrial risk is asset-inventory and access-control failure rather than model behavior—is coherent and well-supported by the FERC audit findings and aerospace c
Source & Claim Verification	Qwen · hosted fallback (qwen-2.5-72b-instruct)	cleared. Most factual claims are supported by citations, but a few lines lack specific references, such as the general statements about aerospace and power grid industries.
Regulatory & Framework Fidelity	Mistral	cleared. The briefing accurately aligns with NERC CIP, IEC 62443, and DO-178C frameworks but omits explicit mapping to ISO 42001, EU AI Act, or NIST AI RMF for industrial AI access risks.
Technical Accuracy	Llama	cleared. The draft demonstrates high technical accuracy in its application of aerospace (ARP/DO), energy (FERC CIP), and industrial (IEC 62443/ISO 26262) regulatory frameworks.
Bias, Balance & Hype Control	Gemini	cleared. The briefing effectively avoids vendor hype and presents a well-supported counterargument to the common focus on AI accuracy, instead highlighting access and inventory as the primary risk.
Novelty & Non-Duplication	Grok	cleared. The aerospace-vs-grid juxtaposition and ‘access/inventory not accuracy’ reframe is a genuine editorial synthesis rather than a wire rewrite, though both source clusters are already circulating and the
Validation	DeepSeek	cleared. The briefing’s central claim about unregistered AI proliferation creating a primary risk of access, not accuracy, is validated by audit findings and industry acquisitions demonstrating AI’s ungoverned

Sources cited: 17. Validation challenges: 0. Review cost: about $0.04. Learn how these briefings are written and verified.