Granite 4.1 LLMON previews use special-token markup to resist prompt injection

Two experimental 3B and 8B models separate instructions from data with reserved tokens, and their developers report sharply lower prompt-injection success on static benchmarks.

AIPublished
What happened
Researchers released granite-4.1-3b-LLMON-preview and granite-4.1-8b-LLMON-preview, two models trained on a new LLM-native markup language called LLMON.
What’s new
LLMON marks each prompt component with reserved special-token delimiters and binds instructions to specific data blocks, rather than relying on JSON or flat text.
Why it matters
The developers report that this separation makes models largely ignore injected instructions hidden in documents, tool outputs or user data.

A team of researchers has published two experimental models on Hugging Face, granite-4.1-3b-LLMON-preview and granite-4.1-8b-LLMON-preview. Both are built from Granite 4.1 base checkpoints and trained on LLMON (LLM Object Notation), a markup language the team designed so models can tell the parts of a prompt apart.

LLMON resembles XML, but its structural markers are reserved tokens in the tokenizer rather than ordinary characters such as brackets or quotation marks. Each piece of a prompt becomes its own block with a short ID: an instruction, a data item, or a requirement. A separate execution block then states which instruction should run on which data. The authors argue that JSON suits parsers rather than language models, because its structural symbols can also appear in normal text.

The main claimed benefit is resistance to indirect prompt injection, where malicious commands are hidden in retrieved content. On the OPI benchmark, the authors say the LLMON models avoided 47% more attacks than the best frontier model they tested and 95% more than the best open-source one. They also report general performance on par with the Granite 4.1 Instruct models.

The developers list several caveats. Their benchmarks cover only static injections, and they say they are still working on adaptive attacks spread over many chat turns. They describe the protection as "a learned behavior, not an enforced guarantee." They also say the models are not intended for production use.

The models ship with a chat template that turns named message parts into LLMON blocks. A companion toolkit converts JSON to LLMON and back without loss.