IBM releases Granite 4.1 with LLMON markup against hidden injections
Prepared by Sintelligent with the help of SI. Published in SI: . Updated: .
Explicitly separating commands from data gives small open models a way to ignore hidden instructions in documents without a separate external filter.
IBM has introduced the experimental granite-4.1-3b-LLMON-preview and granite-4.1-8b-LLMON-preview models. LLMON marks instructions, data and constraints as separate blocks using reserved special tokens, while an execution block explicitly links a command to the relevant data. This lets the model distinguish commands from document content and tool outputs. Lossless JSON conversion tools and examples for adapting the markup to new tasks have also been released. In IBM's tests of static indirect injections, LLMON models avoided 47% more attacks than the best frontier model tested and 95% more than the best open model, while retaining the general capabilities of Granite 4.1 Instruct. The company emphasizes that robustness is learned behavior, not a guarantee, and that adaptive multi-turn attacks require further work. The models are intended for experimentation, not production use.
Sources 1