AI is coming to OT defense. Compliance has to keep up.

In short
AI can take over the clerical core of nuclear OT cyber compliance: matching assets to known vulnerabilities, drafting NEI 08-09 control assessments, answering questions across the compliance record, and flagging assessments that a configuration change has made stale. It does not replace the engineer who signs each assessment. For AI output to survive an NRC inspection, four guardrails matter: it runs inside the plant boundary with no internet connection, a qualified engineer approves every conclusion, every output traces back to the records it used, and the AI never makes autonomous changes to plant systems or records.
This month, Anthropic and eleven security firms launched a program to bring frontier AI to the defense of power grids, water systems, and industrial plants. It is a serious signal: AI-assisted OT security is moving from pilot projects to mainstream offerings.
Most of that conversation is about finding threats faster. Far less of it is about the work that consumes most of a regulated operator’s cyber budget: proving, on paper and to an inspector, that every critical digital asset is identified, assessed and protected. In U.S. nuclear power, that proof is not optional, and AI that cannot survive an NRC inspection is of little use.
Why is nuclear compliance different from other OT cybersecurity?
Because the hard part isn’t detecting threats. It’s proving, with traceable evidence, that every one of a plant’s critical digital assets has been assessed against a required control catalog, in a form that still holds up when an inspector asks about it years later.
Every operating U.S. reactor runs a cyber security program under 10 CFR 73.54, typically implemented through NEI 08-09 or Regulatory Guide 5.71. A single site can carry hundreds of critical digital assets, each needing a documented assessment against a long catalog of technical, operational and management controls.
The burden is not detection. It is evidence. Assessments must be repeatable, traceable to the asset’s actual configuration, and defensible years later. Much of this work is still done in spreadsheets and Word templates, and a full site assessment can take hundreds of engineer-hours.
That makes compliance a natural place for AI to help, and also one of the places where careless AI does the most damage. An assessment conclusion no one can trace back to its basis is a liability, not a time saver.
Where AI genuinely helps compliance teams
Used well, AI removes the clerical load from compliance engineers, so they spend their time on judgment. Four tasks stand out:
- Matching assets to vulnerabilities: mapping each asset’s vendor, model and firmware to known vulnerabilities is tedious by hand. AI can propose matches for an engineer to confirm.
- Drafting control assessments: given an asset configuration and plant procedures, AI can draft a first-pass assessment of each control, with the evidence it relied on cited.
- Answering questions across the record: “which assets still rely on a compensating control for removable media?” AI can answer from the plant’s own records instead of a manual search.
- Keeping assessments current: when a configuration changes, AI can flag which assessments are now out of date, so the program stays continuous.
None of these replaces the engineer who signs the assessment. Each one shortens the path to that signature.
How an AI-assisted assessment moves through the record
| Step | Stage | Description |
|---|---|---|
| 1 | Plant records | Asset inventory, configurations, procedures |
| 2 | AI drafts | Vulnerability matches & control assessments, cited |
| 3 | Engineer reviews | Accepts, edits or rejects each finding |
| 4 | Approved record | Traceable, signed, inspection-ready |
What are the four guardrails that make AI compliance defensible?
In a regulated plant, how AI is deployed matters as much as what it does. Four guardrails separate a defensible tool from a risky one:
Runs inside the plant boundary
Local hardware, no internet connection. Nothing leaves the site; no outside service sits in the loop.
Humans own every conclusion
AI proposes; a qualified engineer accepts, edits or rejects. Every approval is recorded, with who and when.
Every output is traceable
Each AI-drafted finding cites the specific records it drew on, so a reviewer can check the basis directly.
No autonomous changes
AI reads plant records and drafts documents only. It never changes configurations or touches plant systems.
What does this look like in practice?
At WizNucleus, we built these guardrails into Cyberwiz-Pro 7.0, the latest release of our on-premises compliance platform for 73.54 programs. Its AI features run on a local language model inside the plant’s own network, with no internet connection required. The platform runs on standard servers, with GPU hardware optional for heavier AI tasks.
In practice, an engineer works through an assessment the way they always have, but with drafts already waiting. The system proposes vulnerability matches for each asset, drafts control assessments against the NEI 08-09 catalog with supporting records cited, and answers questions about the program in plain language. The engineer reviews, corrects, and approves. Nothing becomes part of the record without that approval.
One U.S. nuclear plant is now running Cyberwiz-Pro 7.0 with these capabilities on a test server, and a multi-site utility fleet is evaluating them. Early estimates point to a 10 to 25 percent reduction in assessment effort.
Five questions to ask any AI compliance vendor
Nuclear operators will hear more AI pitches in the coming year, including from the large OT platforms. Before adopting any of them for regulated work, ask:
- Where does the model run, and does any plant data ever leave the site?
- Can every AI-generated finding be traced to the specific records it used?
- Who approves AI output, and how is that approval recorded?
- Can the AI change anything, or does it only read and draft?
- If the AI feature is turned off, does the program still run on the same records?
A vendor that answers all five clearly is offering a tool an inspector can live with. One that hesitates is offering a demo.
The bar is higher, and that is the point
AI will reshape how critical infrastructure is defended. In nuclear power, it will also reshape how that defense is proven. The winners will not be the tools with the most AI. They will be the ones that make an engineer faster without making an inspector nervous.
Frequently asked questions
What is 10 CFR 73.54?
10 CFR 73.54 is the U.S. Nuclear Regulatory Commission rule requiring every operating nuclear power reactor to maintain a cyber security program that protects critical digital assets from compromise.
What are NEI 08-09 and Regulatory Guide 5.71?
NEI 08-09 is the nuclear industry’s template cyber security plan, and Regulatory Guide 5.71 is the NRC’s guidance for implementing 10 CFR 73.54. Together they define the control catalog nuclear plants use to assess their critical digital assets.
Can AI be used in NRC-regulated nuclear cyber compliance programs?
Yes, when it is deployed as a drafting and research aid rather than a decision-maker. A qualified engineer must still review, approve and sign every control assessment, and the record must show who approved what and when.
Does AI replace nuclear compliance engineers?
No. AI drafts vulnerability matches, control assessments and answers questions about the record, but a qualified engineer accepts, edits or rejects every AI-generated conclusion before it becomes part of the compliance record.
What is Cyberwiz-Pro 7.0?
Cyberwiz-Pro 7.0 is WizNucleus’s on-premises compliance platform for 10 CFR 73.54 programs. Its AI features run on a local language model inside the plant’s own network, requiring no internet connection, with GPU hardware optional.
Does using AI increase NRC inspection risk?
Not when every AI output traces back to the records it drew on and a qualified engineer approves it before it becomes part of the record. Those two guardrails are what let a compliance team answer an inspector’s questions with confidence.


