Self-driving cars face significant testing challenges, constrained by high costs and limited test environments. This necessitates innovative approaches to simulation-based testing to improve deployment and evaluation efficiency. Our study introduces a novel methodology that leverages dangerous real-world road maps sourced from Google Earth as the initial seed for generating driving scenarios. In contrast, traditional approaches use randomly generated seeds/maps to initialize the search process. We systematically adjust road points by evolving these maps to induce out-of-bounds (OOB) errors. Our preliminary results demonstrate a significant improvement in generating failing scenarios, when using real-world maps as seeds compared to random seeds/maps. Specifically, the evolved real-world maps are more likely to be valid (e.g., not self-intersecting) and have a higher incidence of OOB failures. This work opens avenues for further research into optimizing scenario generation for broader applications in autonomous systems testing.