VANETs are highly vulnerable to packet drop and route manipulation attacks due to their inherently distributed nature and rapidly changing topology. Traditional security mechanisms, such as multi-path routing and trust-based forwarding, are limited in practical deployment: the former introduces significant energy overhead, while the latter depends on complex and often unreliable trust evaluations. In this work, we introduce a novel MTD strategy enhanced by AI to dynamically and intelligently adapt routing paths, thereby increasing the uncertainty and cost for adversaries. Specifically, we design a Grid-based Extended Joint Action Learning framework (Grid-eJAL), which leverages MARL to implement an online and adaptive MTD policy tailored for VANETs. Unlike conventional route mutation schemes that rely on static topologies and centralized control, Grid-eJAL supports decentralized, real-time decision-making by enabling vehicles to share learned parameters, accelerating convergence without compromising autonomy. The physical region is partitioned into equal-sized grids, and the AI agent selects the next hop within the optimal grid based on minimum mobility angle and learned defense strategies. The theoretical foundations of Grid-eJAL ensure its convergence. Finally, comprehensive experiments demonstrate its superiority over advanced baseline methods in various adversarial scenarios.