IBM's Granite 4.1 LLMON previews train the model to separate instructions from data, claiming near-zero indirect prompt-injection success. What security teams should test
Reserved special tokens that user content cannot forge, and unique IDs that bind each instruction to the data it may act on: it is the first vendor attempt to make injection resistance a property of the model rather than of the prompt.

Prompt injection has resisted every fix because the model cannot tell where the instruction ends and the data begins. A system prompt says 'treat the following as data'; a retrieved web page says 'ignore the above'; both are text, and the model weighs them the same way. On Wednesday IBM published two experimental models, granite-4.1-3b-llmon-preview and granite-4.1-8b-llmon-preview, trained on a format designed to remove that ambiguity, and reported results that would matter a great deal if they hold up.
What LLMON is
LLM Object Notation, pronounced 'lemon', is an XML-like markup for prompts. Two features do the work. First, the structural delimiters are reserved special tokens in the vocabulary, not ordinary characters, so nothing in user or retrieved content can produce an opening or closing tag; the classic injection trick of mimicking a system marker is impossible at the tokeniser level. Second, tags carry semantic roles, instructions, data, requirements, and unique IDs bind each instruction to the specific data it is allowed to act on, in the way a function's arguments are bound to its parameters. Content that is not referenced by the execution block is meant to be ignored. Existing JSON converts to LLMON and back without loss, and IBM says the format complements chat templates rather than replacing them.
The numbers
On OPI, a benchmark of indirect prompt injections hidden in retrieved content, IBM reports that attack success for the LLMON models drops to nearly zero, while the non-LLMON baselines it tested remain highly susceptible. Expressed as robustness, the LLMON models avoid 47% more attacks than the best frontier model IBM tested and 95% more than the best open-source model. Similar trends are reported on the ASB and StruQ benchmarks. On standard capability benchmarks the previews perform on par with the Granite 4.1 Instruct models, so the resistance does not appear to cost utility. The exact per-model figures are in charts rather than text, and IBM has not named the frontier model it compared against.
The caveats IBM states
This is an experimental preview for evaluation, not production. The injection resistance is 'a learned behavior, not an enforced guarantee': the model has been trained to ignore unreferenced instructions, but nothing in the architecture prevents it from obeying one. The benchmarks cover static indirect injections; adaptive attacks that unfold over multiple turns are still being studied. And the rule that unreferenced content is ignored is itself learned. Those are the honest limits of the approach, and they are the same limits that apply to every other training-based defence.
Why it still matters
The difference from previous defences is where the boundary lives. Guardrail classifiers sit outside the model and can be evaded; system-prompt instructions compete with injected text on equal footing. Reserved-token delimiters move part of the boundary into the tokeniser, where text cannot reach it, and the binding of instructions to data gives the model a reason to ignore an instruction rather than a request to. If the learned behaviour proves robust, a 3B model that can be run on-premises with a measurable injection rate is exactly what regulated deployments of RAG and agents have lacked.
What to test
Take your own indirect-injection set, the payloads that have actually worked against your retrieval pipeline, convert the prompts to LLMON with IBM's toolkit and measure attack success on both previews against your current model. Then try to break the binding: instructions that reference the wrong data ID, injected text that claims an ID, and multi-turn sequences that build up context before the payload. Fine-tune the 3B model on a small task and re-run the injection set to see whether the resistance survives. Report what you find in IBM's GitHub Discussions; the company has asked for it, and the field needs the adaptive-attack results more than another benchmark.
The story so far
- Nov 14, 2026 OpenAI disbands its standalone robotics team for the second time, betting GPT-6 Astra's 74.8% simulation score makes a s…
- Nov 14, 2026 llama.cpp 1.0.0 after 3.5 years: semantic versioning, 35 deprecated functions gone, FP8 KV cache. What the stability pro…
- Nov 13, 2026 Inkling 2 Small's weights are out: 394 GB, Apache 2.0, 500 GB of VRAM to run, and no released serving stack supports it …
- Nov 13, 2026 Gemini 4 Hydrogen: $1.50 / $7.50 flat across a 1M context, 2.5x faster than 3.8 Flash, and the end of the Flash brand
Key terms in this story
- Indirect prompt injection An attack in which malicious instructions are hidden in content the model reads (a web page, email, document or tool result) rather than typed by the user.
- Reserved special token A vocabulary entry a model's tokeniser treats as a control symbol rather than text, so no user or document content can produce it.
- Prompt injection An attack in which text the model reads, such as a web page or email, contains instructions that hijack its behaviour.
- Frontier model A model at or near the state of the art at the time, typically the newest top tier from a major lab.
- System prompt Standing instructions a developer gives a model that apply to every conversation, such as persona, rules and output format.
More in Security
All →
Java4Shell (CVE-2026-114938): an out-of-bounds write in the Java runtime's Base64 decoder, patched Monday, no exploitation seen yet. What to patch, in what order, and how to prove it
The bug is in the JVM, not a library, and it runs before your code gets a vote; that makes it an inventory problem first and a patching problem second…

Microsoft's agent-framework package was poisoned on PyPI for 45 minutes with a stealer that ran at every Python start-up. Two versions, one webhook, at most 50 victims, and a checklist for this morning
The attack did not need you to import the package. A .pth file in site-packages ran the stealer whenever any Python interpreter launched, which is why…

Google's GTIG finds 'exploit_agent', the first framework that turns a vendor patch into a live exploit with no human in the loop. Zero-days hit 22 a month in August. What to change in your patch and detection plan
Google's fourth-quarter AI threat report moves the attacker's clock to the moment a patch is published, documents a backdoor delivered through a codin…
Get the brief in your inbox
One email every morning: the stories that mattered in AI, each with a one-line "why it matters" for your role. Free, no ads, unsubscribe any time.
Published daily by AI Briefing. We never share your address.