Securing AI: What We’ve Learned and What Comes Next