ModelsNews
IBM LLMON: Granite models built to resist prompt injection
IBM has released two experimental Granite 4.1 models trained on a new markup language, LLMON. The idea is to separate instructions from data with reserved special tokens. On the OPI benchmark almost every embedded injection fails, while general performance holds.
· 3 min read

What LLMON is
LLMON (LLM Object Notation, pronounced "lemon") is a markup language designed for LLMs by IBM's Granite team (Hugging Face). It puts each part of the context, such as instructions, data, requirements and tool responses, into its own tagged block. The delimiters are reserved special tokens rather than characters, so structure and content use separate parts of the vocabulary.
Every block gets a unique ID. At the end of a turn, an exec block states which instruction runs on which data, under which requirements, and where the answer goes. Anything not bound into the exec block is material to read, not a command to follow.
Why not JSON?
IBM argues that JSON was designed for parsers, not causal language models. To a model, a quotation mark inside a document and one that closes a field are the same token. Escaping rules are only statistical patterns, and nesting depth must be tracked by counting brackets, which transformers struggle with over long spans. That gap is also an attack surface, and prompt injection exploits exactly that.
What the numbers show
The OPI benchmark hides injections in the data of classic NLP tasks across 9,000 attacks. Both LLMON previews score robustness close to 1, meaning they carry out almost none of the injected instructions. IBM says that is 47 percentage points better than the best closed frontier model it tested, GPT-6-astra, and 95 points better than the best open model. ASB and StruQ show similar trends.
General capability holds: averaged over seven categories, the 3B LLMON scores 69.44 against 68.09 for the Instruct version, and the 8B scores 74.76 against 74.90. The 8B LLMON is weaker in math (79.63 vs 84.09), while both improve in the safety category.

Source: IBM Granite, Hugging Face blog
Where the limits are
IBM stresses that this is learned behavior, not an enforced guarantee. The benchmarks test static indirect injections; work on adaptive attacks that unfold over many turns is ongoing. The models are experimental previews, not for production. For an agent with access to sensitive systems, this can be one layer at most, alongside permissions and isolation.
How to try it
granite-4.1-3b-LLMON-preview and granite-4.1-8b-LLMON-preview are on Hugging Face with their own chat template: just name the parts of a turn (instruction, data, requirement) in the messages list and the template does the rest. The GitHub repo includes lossless JSON to LLMON converters (GitHub), and the Mellea library supports it natively. A LoRA adapter can teach the model your own tag vocabulary. IBM's earlier modular Granite work: Granite Switch 4.1.1.
Key takeaways
- Instructions and data are separated by special tokens and bound by an exec block.
- Near-zero successful indirect injections on OPI, far better than frontier models.
- General performance roughly matches the Instruct models.
- Experimental, learned behavior: it does not replace other defenses.
Produced with the help of AI, reviewed by an editor. How we make our articles
Reviewed and edited by the Neural Daily desk. About


