Safety Assurance of AI-Based Systems: Challenges and A Way Forward