IBM has released two experimental Granite models trained on LLMON, a new markup format meant to keep a model's instructions separate from the untrusted content it reads. IBM described the work in a Hugging Face blog post on November 12. LLMON stands for LLM Object Notation and is pronounced "lemon."
The models are granite-4.1-3b-LLMON-preview and granite-4.1-8b-LLMON-preview, derived from Granite 4.1 base models. Their Hugging Face listings show 4B and 9B parameters. IBM calls them "an experimental preview, ready for exploration and evaluation but not for production."
LLMON wraps parts of a prompt in reserved special-token delimiters and uses an execution block that ties each instruction to the data it applies to. Only instructions reached through that block are treated as authoritative. Documents, tool schemas and tool results are read as material rather than followed as commands, and spans the execution block does not reference are ignored. IBM says this is "a learned behavior, not an enforced guarantee."
On the OPI prompt-injection benchmark, IBM reports that the LLMON models avoid 47% more attacks than the best frontier model it tested and 95% more than the best open-source model. It reports similar trends on the ASB and StruQ benchmarks. These benchmarks test static indirect injections, and IBM says it is still working on adaptive attacks that unfold over many chat turns. It says general performance is on par with Granite 4.1 Instruct models.
Prompt injection is one of the main risks for agents that read email, web pages or tool output. LLMON tackles it by having the model separate trusted instructions from untrusted content, rather than relying on filters outside the model. Developers can test the preview models now, but IBM is clear they are not ready for production.