Few industries hold as much technical documentation as nuclear power. Licensing submittals, design bases, procedures, condition reports, maintenance histories – all of these go back decades at most sites, and a large share of that record started out on paper.
AI in nuclear energy now needs to be applied to that record. National laboratories and operating utilities have spent years testing machine learning and language models against real plant data, and some of those tools have moved past the pilot stage.
The regulator is paying attention, too. The Nuclear Regulatory Commission (NRC) developed its AI Strategic Plan to make sure the agency is ready to review industry uses of AI in regulated activities.
So which AI applications are proving their worth, and what decides where each one is allowed to run?
Just getting started with AI? Check the security essentials first: Setting AI Infrastructure Security Foundations in Regulated Environments
Where is AI in Nuclear Energy Being Applied?
The applications with the strongest track record help people work through large volumes of plant information faster. Much of the published work comes from Department of Energy national laboratories partnering with operating utilities.
Predictive Maintenance
Maintenance is one of the largest operating costs at a nuclear plant. Idaho National Laboratory (INL) researchers note in an extensive study that the U.S. fleet has historically relied on labor-intensive, time-consuming predictive maintenance programs, and that AI and machine learning can simplify complex problems such as diagnosing equipment degradation.
The study regarding INL’s work with Public Service Enterprise Group (PSEG) Nuclear is one of the most detailed public examples. The pilot focused on the circulating water system, an important non-safety-related system that affects the plant’s power generation capability. Over several years, the project produced:
- Fault signatures built from heterogeneous circulating water system data and fault modes at both the Salem and Hope Creek sites.
- Machine learning diagnostic models to automatically detect potential circulating water pump degradation.
- A visualization application called VIPER, designed so that users who may be unfamiliar with machine learning can work with its results.
- A later phase that covered integrating large language models into VIPER to help explain its recommendations.
Corrective Action Program Analysis
The corrective action program (CAP) is how a plant finds and fixes problems. Employees have both the opportunity and the responsibility to identify issues and conditions at the plant that may affect the function of plant systems. Each issue becomes a condition report, and a busy site can generate thousands of them a year.
Screening and trending those reports takes significant staff time. AI is helping in two ways:
- Screening: INL’s MIRACLE tool uses machine learning to provide a confidence score on whether a condition report describes a condition adverse to quality.
- Trending: INL researchers used natural language processing and machine learning to mine hundreds of thousands of condition reports from multiple utilities, producing a set of topics that can describe any report.
Searching Licensing and Design Documentation
Diablo Canyon in California offers the clearest public example of AI applied to document search. The plant’s databases hold about 2 billion pages of documents, and a PG&E vice president estimated that staff spend around 15,000 hours a year just searching for documents.
General-purpose tools fell short; the team tested commercial AI models, and they struggled to work properly with nuclear terminology. Atomic Canyon worked with Oak Ridge National Laboratory to build nuclear-specific models, which were trained on the NRC’s ADAMS database – over 3 million documents. They are designed for retrieval, finding the right source material so that a language model has accurate content to work from.
One detail from the deployment stands out from a security point of view: the enterprise version of the platform runs behind a company’s firewall to meet cybersecurity and export control requirements for business-sensitive data.
What These Use Cases Have in Common
Look across these examples, and a pattern appears:
- They work from copies of plant data, such as historical records, work orders, and document archives.
- Their output is advice that a qualified person reviews and acts on.
- None of them sends commands to plant equipment.
- They depend on computing power and data volumes far beyond what a single on-site server can handle.
These traits are of particular note due to the way the NRC’s cyber security rule draws its lines.
Nuclear energy organizations aren’t the only sector that needs to cement security foundations before rolling out AI tools: AI Data Security in Healthcare and Biotech
What 10 CFR 73.54 Requires of Nuclear Cybersecurity Programs
The NRC issued 10 CFR 73.54 in March 2009, requiring operating nuclear power plant licensees and applicants to protect digital computer and communication systems associated with a plant’s safety, security, and emergency preparedness functions from cyberattacks. It is still the foundation of nuclear cybersecurity for the U.S. operating fleet.
What Does the Rule Protect?
The rule requires licensees to provide high assurance that digital computer and communication systems and networks are adequately protected against cyberattacks, up to and including the design basis threat.
The systems in scope are those associated with:
- Safety-related and important-to-safety functions.
- Security functions.
- Emergency preparedness functions, including offsite communications.
- Support systems and equipment that, if compromised, would adversely affect any of the above.
Licensees must protect these systems from attacks that would modify, destroy, or compromise the integrity or confidentiality of data or software. To do that, a licensee must analyze its digital systems and networks and identify the assets that need protection. These are known as critical digital assets (CDAs.)
How the Rule Becomes a Network Architecture
The regulation sets the outcome. Two guidance documents describe how plants achieve it.
- Regulatory Guide 5.71: Revision 1, released in February 2023, gives licensees guidance on meeting the requirements of 10 CFR 73.54, and clarifies the guidance on defense-in-depth for cybersecurity.
- NEI 08-09: Revision 6 of this Nuclear Energy Institute document provides the licensee cybersecurity plan template that most of the fleet uses.
Both describe a defensive architecture built in levels. The most sensitive systems, including digital safety systems, sit at the highest level, and each level is separated from the next by a boundary device.
The direction of data flow is the defining feature. Boundary control devices enforce one-way communication from Level 4 to Level 3 and from Level 3 to Level 2, and RG 5.71 recommends hardware that enforces one-way communication, such as a data diode.
Revision 1 also widened attention beyond safety systems; it references industry guidance on identifying and protecting CDAs for safety-related, important-to-safety, balance of plant and emergency preparedness functions.
Where Does AI Fit in This Structure?
The defensive architecture governs systems inside the plant’s protected levels. An AI tool that reads a copy of plant data on the business network generally sits outside those levels, as long as it has no path back into them.
The plant’s own cybersecurity plan and CDA assessment make the final call. If an AI tool performs or supports a function within the scope of 73.54, it falls under the rule wherever it runs. Each new use case deserves its own review against the plan.
Most AI Workloads Belong on the Business Network Side
The use cases above fit naturally on the business network. That placement follows from how the tools work, and how plant networks are designed.
Data Already Flows Outward
Plants routinely replicate process data from protected levels to the business network through one-way devices, and engineers already use those copies for reporting and analysis. AI tools can work from the same copies, and the source systems never see an inbound connection.
Outputs Reach the Plant Through People
A maintenance recommendation or a search result has no effect until a person acts on it. That action travels through processes the plant already controls, such as work orders and condition report dispositions. The AI tool needs no connection to plant equipment for its advice to be useful.
Computing Demands Outgrow the Boundary
Training and running modern models takes GPU capacity, frequent software updates and, in many cases, vendor connectivity. The Diablo Canyon models, for example, were developed with an award of 20,000 GPU hours on Oak Ridge’s Frontier supercomputer. The higher defensive levels are designed to keep inbound connections and frequent changes out, so heavy AI workloads are a poor match for them.
Moving Inside Changes a Tool’s Status
An AI tool installed within the protected levels would likely become a CDA itself. It would then carry the full set of controls, assessments and change restrictions that apply to CDAs. Every model update would face the same scrutiny as a change to plant digital systems.
AI Security for Nuclear Power Plants: Controls That Keep Workloads In Their Lane
Putting an AI workload on the business network is the starting point. A small set of cybersecurity controls keeps it there and makes sure it handles plant information in line with regulation requirements.
Hardware-Enforced One-Way Data Flow
A data diode protects the boundary only if it is the sole path across it. Sandia researchers studying advanced reactor designs have noted that wireless communications at Level 4 could allow an adversary to bypass the data diode. The same concern applies to vendor remote access and to maintenance laptops that move between networks.
Before an AI project goes live, map every route plant data takes to reach it. The replicated data store that feeds the AI tool should receive data through the diode, and through no other path.
No Automated Path Back Into the Plant
Many AI platforms offer integrations that write results directly into other systems. Some also include agents that can take actions on a user’s behalf. In a plant setting, these features need close review.
- Route AI recommendations into existing review steps, such as maintenance planning meetings or CAP screening committees.
- Require a named person to approve any record an AI tool creates in work management or CAP systems.
- Disable agent features that can start connections or make changes outside the tool’s own environment.
- Confirm that no integration reaches a system located within the defensive levels.
Identity and Access Management
Who can see the data matters as much as where it is stored, and the cloud adds a further consideration. The Commerce Department has stated that when data or software is uploaded to the cloud, the customer is the exporter and carries responsibility for how that data is stored and who can access it.
In practice, this means:
- Limit access to export-controlled material to verified U.S. persons, using cloud environments designed to enforce that restriction.
- Require multifactor authentication on every AI service and data store.
- Grant access by role, and review it whenever someone changes jobs or leaves.
- Log privileged access and review those logs on a set schedule.
Data Handling Rules
Safeguards Information (SGI) needs its own rule. Anyone who produces, receives, or acquires SGI is subject to the requirements and sanctions of 10 CFR 73.21, whether or not their organization is regulated by the NRC. Those requirements include controls for processing SGI on electronic systems. The safest course is to keep SGI out of AI pipelines entirely, including training sets, search indexes and prompts.
For other sensitive data:
- Encrypt data at rest and in transit, using keys the organization controls.
- Choose cloud regions and tenancy types that match the classification of the data.
Model and Pipeline Governance
AI creates new kinds of data assets. A trained model or a prompt log can all hold information drawn from source documents.
- Keep an inventory of every model, its training data and where it is hosted.
- Treat each model and index as carrying the highest classification of the data used to build it.
- Configure search tools to honor source document permissions, so users only see records they could open directly.
- Retain prompts and outputs in logs the security team can review.
The NIST AI Risk Management Framework provides a useful structure for organizing this governance work.
Vendor and Supply Chain Considerations
Put these questions to any AI vendor before signing:
- Where are the models hosted, and which subprocessors handle customer data?
- Is customer data used to train models shared with other customers?
- How does the platform support U.S.-persons restrictions and export control obligations?
- How much notice will the vendor give before changing models or hosting locations?
A Framework for Classifying Data and Workloads
The controls above depend on knowing what each dataset and workload is. A shared classification framework gives IT, cybersecurity, and licensing teams one consistent way to make that call before anything moves to the cloud.
Step 1: Classify the Data
Sort data into tiers based on where it comes from and which rules govern it.
| Tier | Examples | Cloud Posture |
| Tier 1: Inside the boundary | Live CDA data, control system configurations, network data from Levels 3 and 4 | Stays on site. Only replicated copies leave, through one-way devices |
| Tier 2: Protected by regulation | SGI, Part 810 controlled technology, vendor proprietary design data | SGI stays off AI platforms. Other Tier 2 data goes only to restricted environments with U.S.-persons access |
| Tier 3: Business sensitive | Condition reports, work orders, procedures, non-public licensing correspondence, replicated historical data | Commercial cloud with strong identity management, encryption, and logging |
| Tier 4: Public | Public ADAMS documents, published NRC guidance | Lowest restriction, with access still logged |
Tier 2 also covers Unclassified Controlled Nuclear Information, which is protected under Section 148 of the Atomic Energy Act. When a dataset mixes tiers, treat all of it at the highest tier until the data is separated.
Step 2: Classify the Workload
The same data can feed workloads with very different risk profiles. Ask these questions of every proposed use case:
- Where does each input come from, and what tier is it?
- Which direction does data flow, and is there any return path?
- Does the output influence a plant decision, and through which process?
- Who can reach the model, its prompts and its outputs?
- Where is the model trained and hosted, and what does it keep after a session ends?
A workload starts at the tier of its highest input. It moves to a stricter tier if its output supports a decision covered by the plant’s cybersecurity plan.
NOTE: Match each workload to the right environment. Classification is only useful once it drives a decision. These two steps turn each tier into a deployment choice and keep that choice current.
Step 3: Choose a Deployment Pattern
Each tier maps to one or more deployment patterns:
- On premises only: Tier 1 data, and any workload that would touch SGI.
- Restricted government cloud region: Tier 2 data cleared for cloud use, with access limited to U.S. persons.
- Commercial cloud with strong controls: Tier 3 workloads, such as CAP screening on condition report copies or predictive maintenance on replicated historical data.
- Software as a Service (SaaS): Tier 3 and Tier 4 workloads, once vendor due diligence is complete.
A single use case can span more than one pattern. A document search tool might index public records in a commercial service, for example, while a separate on-premises index holds controlled engineering files.
Step 4: Record and Revisit the Decision
Write down each decision so it can be explained and defended later. A useful record includes:
- The data tiers and workload classification.
- The deployment pattern and the controls applied.
- The business owner and the security owner.
- The rationale and the next review date.
Link these records to existing change management. Reclassify a workload when any of the following happens:
- A new data source is connected.
- The vendor adds a feature such as agent capabilities.
- The hosting location or a subprocessor changes.
- The plant’s CDA list or Cyber Security Plan is updated.
Put the Framework to Good Use
A good first project is a Tier 3 or Tier 4 use case with a clear owner and a measurable payoff. It lets the team test the framework on low-risk data and refine the questions before harder cases come along.
The records produced along the way serve a second purpose: when inspectors, auditors or the board ask how an AI tool was approved, the answer is already documented.
SMS has years of experience partnering with organizations in the nuclear energy sector. Our team advises clients implementing AI under strict regulatory obligations, so we understand the boundaries your programs are built around.
If you’re looking at AI tools, we can show you how to classify your first AI workload. Reach out to our team about where it belongs, and how to secure it.