Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

DriveSafe: A Hierarchical Risk Taxonomy for Safety-Critical LLM-Based Driving Assistants

Abhishek Kumar, Riya Tapwal, Carsten Maple

TL;DR

This work tackles the gap between general-purpose LLM safety frameworks and the domain-specific risks of driving assistants. It introduces DriveSafe, a four-level hierarchical taxonomy with 129 atomic risk types, formally defined by the relation $\mathbb{T} \subseteq \mathcal{D} \times \mathcal{C} \times \mathcal{F} \times \mathcal{R}$ and rooted in real-world regulations. The taxonomy is expert-validated and paired with scenario–prompt data to enable fine-grained safety analysis, with empiricalRefusal assessments across six LLMs revealing substantial gaps in general safety alignment for driving contexts. Overall, DriveSafe provides a practical, domain-aware framework to analyze, simulate, and mitigate safety-critical risks in LLM-based driving assistants, supporting regulatory compliance and safer real-world deployment.

Abstract

Large Language Models (LLMs) are increasingly integrated into vehicle-based digital assistants, where unsafe, ambiguous, or legally incorrect responses can lead to serious safety, ethical, and regulatory consequences. Despite growing interest in LLM safety, existing taxonomies and evaluation frameworks remain largely general-purpose and fail to capture the domain-specific risks inherent to real-world driving scenarios. In this paper, we introduce DriveSafe, a hierarchical, four-level risk taxonomy designed to systematically characterize safety-critical failure modes of LLM-based driving assistants. The taxonomy comprises 129 fine-grained atomic risk categories spanning technical, legal, societal, and ethical dimensions, grounded in real-world driving regulations and safety principles and reviewed by domain experts. To validate the safety relevance and realism of the constructed prompts, we evaluate their refusal behavior across six widely deployed LLMs. Our analysis shows that the evaluated models often fail to appropriately refuse unsafe or non-compliant driving-related queries, underscoring the limitations of general-purpose safety alignment in driving contexts.

DriveSafe: A Hierarchical Risk Taxonomy for Safety-Critical LLM-Based Driving Assistants

TL;DR

This work tackles the gap between general-purpose LLM safety frameworks and the domain-specific risks of driving assistants. It introduces DriveSafe, a four-level hierarchical taxonomy with 129 atomic risk types, formally defined by the relation and rooted in real-world regulations. The taxonomy is expert-validated and paired with scenario–prompt data to enable fine-grained safety analysis, with empiricalRefusal assessments across six LLMs revealing substantial gaps in general safety alignment for driving contexts. Overall, DriveSafe provides a practical, domain-aware framework to analyze, simulate, and mitigate safety-critical risks in LLM-based driving assistants, supporting regulatory compliance and safer real-world deployment.

Abstract

Large Language Models (LLMs) are increasingly integrated into vehicle-based digital assistants, where unsafe, ambiguous, or legally incorrect responses can lead to serious safety, ethical, and regulatory consequences. Despite growing interest in LLM safety, existing taxonomies and evaluation frameworks remain largely general-purpose and fail to capture the domain-specific risks inherent to real-world driving scenarios. In this paper, we introduce DriveSafe, a hierarchical, four-level risk taxonomy designed to systematically characterize safety-critical failure modes of LLM-based driving assistants. The taxonomy comprises 129 fine-grained atomic risk categories spanning technical, legal, societal, and ethical dimensions, grounded in real-world driving regulations and safety principles and reviewed by domain experts. To validate the safety relevance and realism of the constructed prompts, we evaluate their refusal behavior across six widely deployed LLMs. Our analysis shows that the evaluated models often fail to appropriately refuse unsafe or non-compliant driving-related queries, underscoring the limitations of general-purpose safety alignment in driving contexts.
Paper Structure (24 sections, 4 equations, 3 figures, 2 tables)

This paper contains 24 sections, 4 equations, 3 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Refusal Behavior in LLMs
  4. Safety Evaluation Benchmarks and Taxonomies
  5. Safety and Alignment in High-Stakes Domains
  6. Taxonomy Design
  7. Hierarchical Structure
  8. Level 1: Risk Domains ($\mathcal{D}$).
  9. Level 2: Risk Categories ($\mathcal{C}$).
  10. Level 3: Failure Modes ($\mathcal{F}$).
  11. Level 4: Atomic Risk Types ($\mathcal{R}$).
  12. Expert Review
  13. Validation
  14. Limitations
  15. Conclusion
  16. ...and 9 more sections

Figures (3)

  • Figure 1: Hierarchical Risk Taxonomy for Driving AI Systems. The taxonomy is divided into four high-level domains: Techni- cal, Business, Societal and Ethical Risks, which further decompose into categories, failure modes, and specific risk types.
  • Figure 2: Illustrative example showing how a high-level societal risk is refined through the taxonomy into an atomic risk, and subsequently operationalized as a driving scenario and user prompt.
  • Figure 3: Hierarchical risk taxonomy for LLM-based driving assistants. The taxonomy consists of four top-level domains and decomposes into categories, failure modes, and atomic risk types, yielding 129 leaf-level risks.