PNNL

Abstract

This guide is for experienced security assessors who want to use AI—especially large language models (LLMs)—to make security assessments faster, clearer, and more consistent while maintaining strict nuclear security standards. Used correctly, LLMs can help structure complex assessments, generate assessment checklists and question sets, organize notes, and draft non-sensitive report sections and summaries. They can improve documentation quality and support communication with both technical and non-technical stakeholders. At the same time, LLMs introduce critical risks: possible exposure of sensitive information, overreliance on an imperfect tool, and potential dual-use concerns. This guide therefore emphasizes three core principles: 1. Use LLMs as assistants, not decision-makers. 2. Protect sensitive information. 3. Verify all outputs. The guide provides discussion on where LLMs are valuable and where they are not when conducting security assessments. It provides example prompts, takes readers through a sample assessment on a realistic sensitive system, and concludes with next steps that could be taken to further the value of using AI in security assessments.