---
title: IBM releases Granite 4.1 3B/8B LLMON previews that mark instructions, data and requirements with reserved tokens to resist indirect prompt injection | ctxsec
url: https://ctxsec.com/items/ibm-granite-4-1-llmon-preview-instruction-data-separation-prompt-injection-nov-12
description: "LLMON (LLM Object Notation) wraps each prompt component in tags delimited by reserved special tokens, and an exec block binds the one instruction to run to the data it acts on; content reached as input (documents, tool results) is treated as material to read, not commands. IBM says the previews carried out far fewer injected instructions on the OPI benchmark than the frontier and open models it tested (47% more attacks avoided than the best frontier model) with similar trends on ASB and StruQ, while matching Granite 4.1 Instruct on general benchmarks. IBM notes this is learned behaviour, not an enforced guarantee, that it targets static injections rather than adaptive multi-turn ones, and that the models are not for production."
status: "official (IBM Granite blog on Hugging Face, Nov 12, 2026); vendor-reported benchmark results; experimental preview"
published: 2026-11-13
---

> LLMON (LLM Object Notation) wraps each prompt component in tags delimited by reserved special tokens, and an exec block binds the one instruction to run to the data it acts on; content reached as input (documents, tool results) is treated as material to read, not commands. IBM says the previews carried out far fewer injected instructions on the OPI benchmark than the frontier and open models it tested (47% more attacks avoided than the best frontier model) with similar trends on ASB and StruQ, while matching Granite 4.1 Instruct on general benchmarks. IBM notes this is learned behaviour, not an enforced guarantee, that it targets static injections rather than adaptive multi-turn ones, and that the models are not for production.

[← AI](/ai) · [markdown](/items/ibm-granite-4-1-llmon-preview-instruction-data-separation-prompt-injection-nov-12.md)

official (IBM Granite blog on Hugging Face, Nov 12, 2026); vendor-reported benchmark results; experimental preview

2026-11-13 7:40 AM ET

### IBM releases Granite 4.1 3B/8B LLMON previews that mark instructions, data and requirements with reserved tokens to resist indirect prompt injection

- **Use it as one layer:** keep tool permissions, human approval and output checks; IBM itself says the separation is not a guarantee.
- Models, toolkit and converters are open: `granite-4.1-3b-llmon-preview`, `granite-4.1-8b-llmon-preview`, github.com/generative-computing/llmon.

Sources: [IBM Granite (Hugging Face blog): LLMON, now in Granite 4.1 3B and 8B](https://huggingface.co/blog/ibm-granite/llmon-an-llm-native-markup-language-now-in-granite) · [Hugging Face: ibm-granite/granite-4.1-8b-llmon-preview](https://huggingface.co/ibm-granite/granite-4.1-8b-llmon-preview) · [arXiv: Post-Training Large Language Models with LLMON](https://arxiv.org/abs/2610.38120)
