The operation safety of autonomous vehicles (AVs) faces great challenges under the emerging AI model-driven R&D paradigm, where artificial intelligence (AI) safety must embrace the impact of critical, complex, and random scenarios. Multiple measures have been proposed for safety assessment and assurance, and it is necessary to develop a comprehensive strategy to address specific driving tasks under different driving conditions. This study presents a comparative overview of the state-of-the-art safety assurance measures based on safety assessment metrics. The safety assessment metrics are categorized into four groups, i.e., scenario criticality, situation complexity, scene consistency, and vehicle self-safety handling envelope. Based on this embodied risk perception, AVs can take different measures for safety assurance at different risk evolution stages. Three typical operation safety measures, i.e., proactive safety behaviors, reactive safety response, and emergency evasive operation, are conducted to filter risk in a hierarchical manner. A system-level architecture design for safety assurance of AI-driven AVs is presented, in which a dedicated safety monitor unit is designed to capture the safety assessment metrics. The system architecture is also compatible with mainstream end-to-end autonomous driving system (ADS) to expand the safety operational design domain (ODD) boundary, and operation risks are filtered out through hierarchical validation and inhibition. Conclusions and future studies are also highlighted. The comparative overview is expected to assist with accident prevention of AVs.