IBM's Granite LLMON cuts prompt injection attacks to near zero
Two experimental Granite 4.1 previews read a new markup language that separates instructions from data, built to stop prompt injection hidden in retrieved documents.
IBM’s Granite team released two experimental preview models on 12 November: granite-4.1-3b-llmon-preview and granite-4.1-8b-llmon-preview. Both read LLMON, LLM Object Notation, a markup language built to keep trusted instructions and untrusted data visibly separate inside a prompt.
LLMON uses six reserved tokens as delimiters rather than ordinary characters. Each piece of context gets a unique ID, and instructions only apply to data through explicit execution blocks; anything outside a bound block is treated as text to read, not a command to follow. On IBM’s OPI prompt-injection benchmark, the preview models blocked 47% more attacks than the strongest frontier model IBM tested and 95% more than the best open-source model, while matching the standard Granite 4.1 Instruct models on general benchmarks. Code and weights are on GitHub and Hugging Face under Apache 2.0.
Why it matters: if you’re building a RAG pipeline or an agent that reads documents it didn’t author, injected instructions hiding in that data are the main attack surface. LLMON is a structural answer to that rather than another system-prompt patch, and it’s free to inspect now.
The caveat: IBM calls these previews “ready for exploration and evaluation but not for production.” They’re 3B and 8B only, and the protection only applies if your pipeline actually renders content into LLMON’s format rather than plain chat messages.