Welcome Your New AI Teammate: On Safety Analysis by Leashing Large Language Models

This paper presents a systematic prompt engineering pipeline for automating Hazard Analysis and Risk Assessment (HARA) using LLMs. The authors design a series of chained prompts — each handling a specific HARA subtask — that propagate context through the analysis to produce an initial 'version zero' of the safety assessment without human intervention. Key prompt engineering patterns identified include role assignment, context injection, template-constrained output, and task decomposition. Validated against automotive safety standards (ISO 26262), demonstrating that carefully engineered prompts can produce structured, useful safety artifacts that accelerate the work of safety engineers. Relevant to PSM because the prompt pipeline methodology is directly transferable to HAZOP, LOPA, and What-If studies.