IBM announced two preview models, granite-4.1-3b-LLMON-preview and granite-4.1-8b-LLMON-preview, built on its Granite 4.1 base. The models are trained to read and generate LLMON, a markup that tags instructions, data and requirements with special tokens. IBM says the new format cuts indirect prompt‑injection success rates to near zero while keeping overall capability comparable to standard Granite 4.1 Instruct models. The code, a GitHub repo and conversion tools are also released for developers to try.
Why it matters
Developers can test a safer way to separate commands from user data, reducing the risk of hidden prompt attacks in their applications.
