Home Hacking Words as Weapons: What 300K Prompt Injection Attacks Taught Us About AI Security