AI exploits just breached 3 major companies. This isn't theoretical anymore - adversarial attacks are now hitting production systems at scale. The attack vectors likely involved prompt injection, model poisoning, or exploiting API endpoints that interface with LLMs. Classic security assumptions break down when your auth layer can be socially engineered through natural language. Companies running GPT-4 or Claude in customer-facing apps without proper sandboxing are basically running arbitrary code execution with extra steps. The real question: were these zero-days in the models themselves, or just bad implementation security? Either way, this is a wake-up call that AI security can't be an afterthought bolted on post-deployment. Every company integrating LLMs needs red-teaming specifically for AI attack surfaces now.