最新模型
IBM Opens Granite LLMON Preview Models, Using Structured Instruction Binding to Reduce Prompt Injection Risks
On November 12, IBM released experimental 3B and 8B LLMON models based on Granite 4.1, making the relationships between instructions, data, and tool results explicit. Official tests show improved resistance to static indirect prompt injection, but this remains learned behavior, not a security guarantee.

On November 12, IBM made preview LLMON models based on Granite 4.1 3B and 8B available, along with a companion tool library. The release changes both the prompt format and model post-training: starting from Granite 4.1 base checkpoints, the models learn to distinguish text with instruction authority from text that is merely data to process. They are currently intended for experimentation and evaluation. Official announcement
LLMON divides system rules, user instructions, documents, and tool results into addressable blocks. An exec block specifies which instructions to execute, as well as the input and output locations. Structural boundaries use six reserved special tokens, while nested relationships are expressed with tag paths. The key engineering idea is to give the model data relationships already known to the application, reducing the need for the model to infer them from mixed text. Model card
For integration, developers can continue to use apply_chat_template to generate markup with the model’s included template. To distinguish content explicitly, a message can be split into fields such as instruction, data, and requirement. Blocks marked exec: false remain in context but are not referenced by execution blocks. The companion repository also provides JSON conversion tools that preserve structured data as typed, addressable blocks for use in prompting and training workflows. Tool repository
This is especially relevant to RAG and tool-using agents: retrieved documents or tool outputs should be treated as input data even when they contain imperative language. IBM says the models show marked improvement on static indirect prompt injection evaluations such as OPI. However, the company also states that multi-turn adaptive attacks still need improvement, and that the distinction between instruction authority and data is learned model behavior, without an enforcement guarantee. Evaluation details
Teams should still assess capability trade-offs before adoption. The 3B model card reports an MMLU-Pro score of 46.08, below the general Instruct version’s 49.83, showing that similar overall capability does not mean performance is even across tasks. Engineering teams should next validate data boundaries, task quality, and template integration using their own Chinese documents, tool responses, and multi-turn attack cases. Capability evaluation