IBM released LLMON, which binds instructions, data and requirements using reserved delimiter tokens and identifiers, alongside Granite 4.1-based 3B and 8B previews. It reports sharply lower attack success on static indirect prompt-injection benchmarks while maintaining general capability by separating instructions from input material. JSON round-trip conversion tools and Mellea integration are included. Adaptive multi-turn attacks remain under study: the separation is learned, not enforced, and these experimental models are not intended for production.