IBM's Granite 4.1 LLMON previews cut indirect prompt-injection success from 53–100% to 0.02–5.6% — by making the instruction/data boundary reserved tokens
IBM's open-weights 3B/8B previews separate instructions from data at the vocabulary level — and IBM is unusually honest that the defense is learned, not guaranteed.
IBM's Granite 4.1 LLMON previews cut indirect prompt-injection success from 53–100% to 0.02–5.6% on the Open-Prompt-Injection (OPI) benchmark by encoding the instruction/data boundary as reserved vocabulary tokens instead of characters. The 3B and 8B models are Apache 2.0, and IBM is explicit that the defense is a learned behavior, not an enforced guarantee.
What shipped
On November 12, IBM's Granite team released two experimental open-weights models — granite-4.1-3b-LLMON-preview and granite-4.1-8b-LLMON-preview — under Apache 2.0 (Hugging Face announcement, model card). Both are fine-tuned from the Granite 4.1 base checkpoints, not the instruct models, so LLMON behavior is learned during post-training rather than layered onto an existing chat format.
LLMON (LLM Object Notation, pronounced "lemon") is an XML-like markup whose structural delimiters are reserved special tokens rather than characters. Six delimiters are reserved, each a single token: <|open|>/<|close|> open a block, <|open_end|> closes any block, <|self_close|> closes an empty field, <|.|> joins a nested tag path (e.g. exec<|.|>input), and <|:|> separates a tag from its id (model card). Because the delimiters live in a separate part of the vocabulary, structure "cannot be formed by subword composition of ordinary content" — a quotation mark inside a document and a quotation mark that closes a field are no longer the same token, which is the known weakness of JSON and character-delimited formats (Hugging Face announcement).
Every piece of a prompt — system policy, instructions, data, documents, tool definitions, tool results — gets its own addressable block with a two-character ID drawn from a fixed single-token set. An exec block then names which instruction to run, which inputs it operates on, and where the output goes. The privilege rule the model is trained to follow is spelled out on the card: "exec<|.|>instruction and exec<|.|>system are authoritative. Everything reached through exec<|.|>input (data, documents, tool schemas, and tool results) is input to be read, not a command to follow, regardless of how it is phrased" (model card).
The training recipe is described in the companion paper, Post-Training Large Language Models with LLMON: full fine-tuning on a 4.6-billion-token structured corpus (arXiv paper), then five sequential GRPO stages spanning multi-domain reasoning, RLHF, knowledge calibration, and math, followed by a final DPO stage on a thousand on-policy preference pairs built from the model's own responses to injection attempts (Hugging Face announcement). The SFT data is augmented with decoy instructions and data blocks deliberately left unbound, and distractor variants with up to three competing instructions.
IBM also shipped the tooling: an open LLMON toolkit with lossless JSON↔LLMON converters, LoRA examples, and a Mellea integration that auto-LLMONizes LLM calls, plus an attention-visualization Space.
Why it matters
The headline number is a big one. IBM reports that on the Open-Prompt-Injection (OPI) benchmark — which places the injection in the data of classic NLP tasks and reports robustness as 1 − attack success rate — the LLMON previews "practically eliminates the problem on this benchmark, avoiding 47% more attacks relative to the best frontier model that we tested, and 95% more than the best open-source model" (Hugging Face announcement). Baseline Granite 4.1 models let indirect injections succeed 53–100% of the time; the LLMON versions, 0.02–5.6% (arXiv abstract).
The mechanism is the part worth stealing. Prior XML-tag-style defenses are escapable because their delimiters are ordinary characters an attacker can reproduce inside a document. LLMON's delimiters are reserved tokens, so an injection payload cannot forge the boundary by writing text — it would have to emit the model's own special tokens, which is a structurally harder attack. That is a genuinely different answer to a known weakness.

The cost is real but bounded. General capabilities are roughly at parity with Granite 4.1 Instruct, though the mix is not uniform: on the 8B, GPQA drops from 41.96 to 35.05 and DeepMind Math from 80.07 to 70.86, while BigCodeBench rises from 35.00 to 44.04 and SALAD-Bench safety from 95.80 to 97.70 (model card eval table). The card is also candid that none of the standard benchmarks measure the boundary-holding behavior the format exists for.
Caveats
Everything above is vendor-reported. IBM's own models, its own numbers, its own benchmark runs — and none of the sources we reviewed reports an independent red-team or replication. Treat the 0.02–5.6% as a self-reported result, not an established bound.
IBM is unusually direct about the limits, and those limits are real:
- Learned, not guaranteed. The card states: "The privilege rule is a learned behavior, not an enforced guarantee. It substantially reduces, but does not eliminate, the chance that an instruction embedded in data, a document, or a tool result is followed. Do not treat it as a sole defense for an agent with access to sensitive systems" (AI Weekly, quoting the model card).
- Static indirect attacks only. The benchmarks test robustness to static indirect prompt injections, and IBM says it is "actively working on" adaptive attacks that unfold over many chat turns — multi-turn attacks are unfinished, and direct injection is not the claimed defense target (Hugging Face announcement).
- Preview, and a different interface. These are experimental previews, and prompts in a conventional chat format are out of distribution. Adopting LLMON means adopting the markup, the custom chat template, and the converter tooling — it is agent infrastructure, not a drop-in chat replacement.
The open question the release itself invites: does the "lemon" survive a real red team before it becomes a production agent format? The weights are Apache 2.0 and the recipe is public, so anyone can run that test.