AED leverages large language models to automatically discover critical safety vulnerabilities in autonomous driving systems that are both effective and diverse.

• Improves upon traditional reinforcement learning approaches by finding failures where the autonomous vehicle is genuinely responsible
• Identifies a broader spectrum of vulnerability types compared to existing methods
• Provides a framework for comprehensive safety assessment before real-world deployment
• Demonstrates how AI tools can enhance rather than compromise autonomous system security

This research is vital for the security industry as it helps prevent potentially catastrophic autonomous driving failures by systematically identifying weaknesses before vehicles reach public roads, ultimately accelerating safe deployment of autonomous technology.

AED: Automatic Discovery of Effective and Diverse Vulnerabilities for Autonomous Driving Policy with Large Language Models