Key Points
- IBM's Granite team published LLMON, an XML-like markup language built for large language models.
- Two preview models separate executable instructions from documents, tool schemas and tool outputs they read.
- On the OPI benchmark LLMON avoided 47% more attacks than the best frontier model tested.
The latest:
Two experimental Granite models trained to treat instructions inside a dedicated exec block as the only commands they obey were released on 12 November 2026, alongside a new markup language called LLMON. The IBM Granite team wrote on the Hugging Face blog that documents, tool schemas and tool results are read as material, not executed. The authors described it as learned behavior rather than an enforced guarantee.
Details:
- The format: LLMON, short for LLM Object Notation and pronounced lemon, is an XML-like markup language placing context segments inside meaningful tags bounded by special reserved-token delimiters. The authors argued JSON was designed for parsers rather than causal language models, and that separate-token delimiters make nesting and boundaries easier to track.
- How it binds: Unique identifiers and an exec block tie instructions to the specific data they are meant to act on, in the way parameters are passed to a function. According to the post, LLMON enriches chat templates rather than replacing them, letting the user label parts of a turn as instruction, data or requirement.
- The release: IBM published granite-4.1-3b-llmon-preview and granite-4.1-8b-llmon-preview, both derived from Granite 4.1 base models, plus a GitHub repository at generative-computing/llmon carrying tooling, JSON-to-LLMON converters, examples and templates. The post also cited native support in Mellea and an attention-visualization demo hosted on Hugging Face Spaces.
- The training: Training started from Granite 4.1 base checkpoints rather than instruction models. Six delimiters were added to the tokenizer as reserved tokens, the instruction-tuning set was rewritten in LLMON with unbound decoy blocks, and training ran through 5 sequential GRPO stages followed by a DPO stage.
- The preference data: The final DPO stage used 1,000 preference pairs drawn from model responses to attempted injections, the authors wrote. The resulting models are trained to ignore any segment the exec block does not point to, including instructions smuggled into retrieved documents or returned by a tool call.
- The numbers: Robustness was measured as 1 minus attack success rate across the OPI, ASB and StruQ benchmarks. On OPI, LLMON avoided 47% more attacks than the best frontier model tested and 95% more than the best open-source model tested, with attack success rates falling to near zero on some benchmarks.
- General capability: The authors reported general capability remains on par with Granite 4.1 Instruct across standard benchmark categories, meaning the injection resistance did not come at a measured cost elsewhere. Models without LLMON were described as highly vulnerable in the same tests.
- The limits: The evaluations cover static injection only, and the authors said adaptive attacks spread across multiple conversation turns remain work in progress. The post described the release as “an experimental preview, ready for exploration and evaluation but not for production.”
- Extensibility: Users can define their own tags and fine-tune a LoRA adapter on top of the LLMON models, according to the post, which also pointed to two related papers filed as arXiv 2610.38120 and arXiv 2603.22519.
Between the lines:
The framing matters for anyone building agents. Because the models are trained to ignore unreferenced segments rather than mechanically blocked from acting on them, the authors call it learned behavior, not a guarantee — and a near-zero attack success rate on static benchmarks is not the same as safety against the adaptive, multi-turn attempts the team says it is still working on.
What’s next
Watch for results on adaptive multi-turn injection attacks, independent benchmark replications on OPI, ASB and StruQ, and whether IBM moves either preview model from experimental status toward a production release.