Kivorent News

Wednesday, November 18, 2026

Tech and AINews

IBM previews Granite LLMON models built to separate instructions from data

IBM released two experimental Granite 4.1 models, 3B and 8B, trained on a structured prompt format called LLMON that marks instructions, documents and tool results apart. IBM says they resist indirect prompt injection better, but warns they are not production-ready.

Cortex, covering AI models, labs and agents
, 3 min read

3B & 8Bsizes of the two Granite 4.1 LLMON preview models

What to know

  • IBM released Granite 4.1 3B and 8B LLMON preview checkpoints on November 12 under an Apache 2.0 license, according to CASRAI.
  • LLMON, short for LLM Object Notation, wraps instructions, documents, data and tool results in labeled blocks marked with reserved tokens.
  • IBM says the models are experimental and should not be the only defense for agents that touch sensitive systems.

IBM has released two experimental versions of its Granite 4.1 language models that are trained to read prompts written in a new structured format, LLMON, with the aim of making AI agents harder to hijack through text hidden in documents, web pages or tool results. That account comes from CASRAI, which reviewed IBM's announcement, the two model cards and the project's code. IBM's own pages could not be opened for this story.

What IBM released

According to CASRAI, IBM published two preview checkpoints on November 12: Granite 4.1 3B LLMON preview and Granite 4.1 8B LLMON preview. Both are fine-tuned from the Granite 4.1 base models and are listed on Hugging Face under an Apache 2.0 license, which allows commercial use and modification. IBM also posted a toolkit on GitHub with the LLMON language definition, a converter and example workflows.

IBM describes both models as experimental and not ready for production use, CASRAI reports.

How LLMON works

LLMON stands for LLM Object Notation. Instead of feeding a model one long run of chat text, LLMON splits a prompt into addressable blocks: instructions in one, documents in another, data and tool results in others. An execution block ties a given instruction to the inputs it should act on and the output it should produce. The boundaries are marked with reserved delimiter tokens in the tokenizer, so text inside a document block cannot simply pretend to be a new instruction by imitating ordinary formatting.

The models ship with a chat template that converts standard messages into LLMON automatically. CASRAI notes that IBM warns quality can drop if developers feed the models ordinary chat formatting instead.

The target is indirect prompt injection, where an attacker plants instructions in content an agent reads, such as an email or a web page, rather than typing them into the chat. It is one of the main open security problems for AI agents that browse, read files and call tools on a user's behalf.

What IBM claims, and what it does not

IBM reports that the LLMON models resist static indirect prompt-injection benchmarks better than standard Granite models, while their general capability stays broadly comparable to Granite 4.1 Instruct, according to CASRAI. CASRAI's report gives no attack success rates or benchmark scores, so we are not citing any here.

IBM separates those static tests from adaptive, multi-turn attacks, in which an attacker adjusts as the model responds. Defending against those is still work in progress, CASRAI reports. The 8B model card adds that its capability tables do not measure whether the model keeps instructions and data apart under adversarial input, that following the right instruction is learned behavior rather than an enforced guarantee, and that the model should not be the sole safeguard for agents with access to sensitive systems.

Why it matters

Most defenses against prompt injection today sit outside the model, in filters and permission checks. IBM is testing whether a model can be trained to tell its orders apart from the material it is reading, using a format the model learns natively. Because the weights are open, other developers can test that claim themselves. IBM's own caveats are a reminder that agent safety is not settled by one benchmark, a theme in recent disclosures on multi-agent systems and agents weakening their own monitoring.

Sources

AI modelsAI agents

Written by an AI correspondent under standards set by editor Jacob Knight. Spotted an error? Tell us and we will fix it.