Illustration, AI-generated from the story text. Tap a panel to read the story through it.

IBM released experimental previews of Granite 4.1 models trained on LLMON, an XML-like markup whose delimiters are reserved special tokens.

A prompt tags each instruction and each piece of data with an ID, and an execution block binds instructions to the data they may act on, so text arriving inside a document is treated as data.

IBM reports that on the OPI prompt-injection the models avoid about 47% more attacks than the best it tested and about 95% more than the best open model, with general ability on with Granite 4.1 Instruct.

IBM says the models are not for production. The benchmarks use static injections, and the authors are still testing attacks spread over many turns. The separation is learned behavior and the model can still be talked out of it. The blog post calls the models 3B and 8B; their model cards list 4B and 9B.
IBM trains Granite models to tell instructions from data with reserved markup tokens
IBM released experimental previews of Granite 4.1 models trained on LLMON, an XML-like markup whose delimiters are reserved special tokens. A prompt tags each instruction and each piece of data with an ID, and an execution block binds instructions to the data they may act on, so text arriving inside a document is treated as data. IBM reports that on the OPI prompt-injection benchmark the models avoid about 47% more attacks than the best frontier model it tested and about 95% more than the best open model, with general ability on par with Granite 4.1 Instruct.
IBM says the models are not for production. The benchmarks use static injections, and the authors are still testing attacks spread over many turns. The separation is learned behavior and the model can still be talked out of it. The blog post calls the models 3B and 8B; their model cards list 4B and 9B. SourcesA
Sources
This story ran in the Afternoon Brief of November 13th 2026.