IBM’s Granite preview teaches models to separate instructions from data
IBM has released experimental Granite models trained to read explicit boundaries between commands, input material and output requirements. The LLMON format addresses a common route for hostile instructions hidden inside documents or tool responses. The preview includes developer tools and conversion examples, while IBM keeps production use outside its recommendation and describes the boundary as learned behavior.
Artificial Intelligence··Midday
A format becomes preview models
IBM released experimental Granite models on 12 November through Hugging Face, a platform for sharing machine-learning models. They are trained to consume and generate LLMON, a markup language that labels instructions, input data and output requirements separately. The release brings model weights and developer tools to the earlier work on the notation. IBM describes the models as an experimental preview for exploration and evaluation, and explicitly advises against using them in production services.[1], [2]
Linking commands to their inputs
A LLMON execution block identifies which instruction applies to which input and which requirements govern the answer. Reserved delimiters keep the structure distinct from ordinary text, and identifiers connect the relevant blocks. Training includes misleading commands that appear in context without being assigned authority through that execution block. The intended result is that a command hidden in a retrieved document or a tool response is treated as material to read, rather than an instruction to obey.[1]
The boundary remains learned
IBM reports tests against indirect prompt injection, the insertion of hostile commands into data a model reads. Those evaluations are the developer’s own results, and IBM says the privilege boundary is learned behavior rather than an enforced guarantee. IBM’s work on adaptive attacks across repeated exchanges continues. Conversion tools let developers move structured data between LLMON and existing formats, while Mellea integration examples support experimentation. The preview gives developers a way to test the approach while keeping its security limits explicit.[1]
