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

TL;DR

The paper investigates using Large Language Models (LLMs), specifically GPT-4.0, to automate Hazard Analysis & Risk Assessment (HARA) within automotive safety engineering to accelerate SafetyOps in autonomous vehicle development. It proposes a structured, six-step HARA pipeline and a set of prompt-engineering patterns (P1–P4) to propagate task context, generate hazard scenarios, assess severities, formulate safety goals, and produce a CSV HARA table, while requiring ongoing expert review. Using a novel CAEM function as a case study and a model-agnostic approach, the study demonstrates that an initial version-zero HARA can be generated rapidly, enabling engineers to brainstorm and iterate more quickly, albeit with limitations in meeting industrial acceptance criteria and the need for independent evaluation. The findings highlight the practical potential and challenges of integrating LLMs into safety analysis workflows, underscoring the importance of transparency, regulatory alignment (e.g., AI Act), and a human-in-the-loop process, with future work focusing on broader validation, prompt/pipeline refinements, and hybrid approaches that combine rule-based tools with LLM reasoning.

Abstract

DevOps is a necessity in many industries, including the development of Autonomous Vehicles. In those settings, there are iterative activities that reduce the speed of SafetyOps cycles. One of these activities is "Hazard Analysis & Risk Assessment" (HARA), which is an essential step to start the safety requirements specification. As a potential approach to increase the speed of this step in SafetyOps, we have delved into the capabilities of Large Language Models (LLMs). Our objective is to systematically assess their potential for application in the field of safety engineering. To that end, we propose a framework to support a higher degree of automation of HARA with LLMs. Despite our endeavors to automate as much of the process as possible, expert review remains crucial to ensure the validity and correctness of the analysis results, with necessary modifications made accordingly.

Figures (1)

• Figure 1: LLM-based HARA, utilizing a pipeline of subtasks, each managed through a specific prompt. The item definition is imported (top-left), and the HARA results are exported (bottom-right). In the second row of the HARA table, the relationship of each column to the prompts is summarized.