Secure Generative AI for Utilities: Bridging the Air Gap

The Unstoppable Force vs. The Immovable Object

The utility industry is currently caught in a paradox of historic proportions. On one side, there is the unstoppable force of Generative AI — technology that promises to revolutionise how we query network data, automate designs, and predict failures. On the other side is the immovable object of Critical Infrastructure (CI) protection.

Utilities manage the grid, the lifeblood of modern society. The data describing this grid — asset locations, capacities, customer connections — is sensitive. It cannot simply be pasted into a public Large Language Model (LLM) like ChatGPT.

For GIS Managers and CIOs, this creates a state of paralysis. You see the potential of AI to query your Smallworld Geo Network Management (GNM) database (“Show me all transformers installed before 1990 in Zone B”), but you cannot risk that query training a public model. The nightmare scenario is clear: a malicious actor using “adversarial prompting” on a public AI to extract the topology of your substation network.

The solution is not to ban AI. That is a path to obsolescence. The solution is to architect a “Red/Black” separation layer — a middleware approach that allows your Smallworld GNM environment to “speak” to AI without ever whispering its secrets.

Deconstructing the “Red/Black” Architecture

The “Red/Black” concept, traditionally used in cryptography and secure facility engineering (TEMPEST standards), refers to the physical separation of systems.

• Red Side: Systems containing sensitive plaintext information (your Smallworld GNM data).
• Black Side: Systems carrying encrypted or unclassified information (the public internet / AI model).

In the context of MagikDev’s Assistant for Smallworld GNM, we apply this rigorous separation to the data flow between GE Vernova’s GNM and the AI model. This is not just a feature; it is the fundamental architecture of the product.

How the Middleware Shield Works

Most AI implementations fail security audits because they connect the data source directly to the model. They feed the AI a CSV file of asset data and ask it to find patterns. This is the “Data Leakage” vector.

MagikDev’s approach is radically different. We introduce a proprietary middleware layer that acts as a semantic translator.

1. The User Request (Red Side): A user in Smallworld GNM asks, “Highlight all cables with low oil pressure.” This request happens inside your secure firewall.

2. The Translation (The Air Gap): The middleware intercepts this request. Crucially, it does not send the cable data to the AI. Instead, it sends a schema definition — a map of how to ask for cables — to the AI. It essentially says: “I have a collection of objects called ‘cables’ with an attribute ‘oil_pressure’. If a user wanted to find low ones, what query logic would you recommend?”

3. The AI Instruction (Black Side): The AI, understanding the schema but seeing no data, returns a block of Magik code or a query structure: run_query(cable_collection, 'oil_pressure < threshold').

4. Execution (Red Side): The middleware receives the code, sanitises it, and executes it within the secure Smallworld GNM environment. The AI never saw a single cable record; it only saw the logic required to find them.

Why “No Data to AI” is the Only Viable Path

Recent studies on Generative AI risks in critical infrastructure highlight “supply chain attacks” and “data poisoning” as top threats. If your proprietary network topology is used to train a public model, that topology creates a permanent vulnerability.

MagikDev’s Assistant is built on a strict “No Data to AI” protocol.

• Vendor-Agnostic Flexibility: Because the middleware handles the translation, you are not locked into a single AI provider. Whether it’s Gemini, Copilot, or Grok, the security posture remains identical. You can swap the “brain” (the LLM) without changing the “body” (your secure data).
• Local Execution: All “thinking” regarding specific data values happens locally. The AI provides the template for the thought, but your local server fills in the blanks.

The Operational Benefit: From Query to Action

Case Study: The Complex Trace

• The Old Way: A user needs to trace a feeder from the substation to all downstream commercial customers. This requires knowing the specific trace predicate in Magik or navigating a complex UI menu.
• The AI Way: The user types: “Trace feeder 1234 downstream and select all commercial meters.”
• The Mechanism: The AI translates the natural language into the correct Smallworld GNM trace config. The middleware executes the trace. The user gets the result in seconds.

At no point did the list of commercial customers leave the secure environment.

The Future is Hybrid

The GE Vernova roadmap points toward increased automation and “Grid Orchestration.” As you migrate toward GridOS and Smallworld GNM 5.x, the volume of data will outpace human ability to query it manually. AI is inevitable.

By adopting a secure, middleware-driven architecture today, utilities can unlock the efficiency of AI — saving hours on complex queries and quality control rules — without compromising the integrity of the grid. It is possible to have the agility of a startup and the security of a utility. It just requires a little “Magik.”

Learn about the Assistant →