LLMON: IBM's New Markup Language Treats Prompt Injection Like a Type Error

IBM's Granite 4.1 LLMON models use reserved special tokens to separate instructions from data—dropping indirect prompt injection success rates to near-zero while maintaining benchmark parity.