AI Briefing
KOSign in

IBM Releases LLMON Markup Language and Granite 4.1 Preview Models with Enhanced Prompt Injection Robustness

·2026.11.13 04:07

Key point

The experimental preview models use reserved special tokens to separate instructions from data, achieving higher robustness against prompt injections without sacrificing general benchmark performance.

1 / 3

Details

  • IBM Granite team introduces LLMON (LLM Object Notation), an XML-like markup language that uses reserved special tokens to explicitly bind instructions to data payloads.
  • LLMON models demonstrate 47% higher robustness (defined as 1 - attack success rate) against indirect prompt injection attacks compared to the best frontier model tested, and 95% higher robustness than the best open-source model on the OPI benchmark.
  • The approach maintains general benchmark performance equivalent to standard Granite 4.1 Instruct models, ensuring safety improvements do not degrade core capabilities.
  • Training involves post-training on Granite 4.1 base checkpoints with a privilege rule that treats execution block instructions as authorized and input data as read-only.
  • Available as experimental previews (granite-4.1-3b-LLMON-preview, granite-4.1-8b-LLMON-preview) with a lossless JSON-LLMON conversion tool and LoRA support for custom domain vocabularies.
open-sourcemodel-releaseprompt-injectionibmgranitellmonmarkup-language
Read originalShare on Threads

This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.

Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.

Related

RedditNex-N2.6 releases public FP8 weights with 16% to 1.6% repetition reductionHuggingFace BlogIBM releases Granite Speech 5.0 with improved ASR accuracy and 1.6x faster Turbo CTCHuggingFace BlogIBM releases Granite Switch 4.1.1 with updated Speech ASR and Decision adaptersRedditbasal-omni-2.0-mini released with 0.811 Werdykt score and two-way state readingPyTorchKRllama.cpp v1.0.0, First Major Release in Three and a Half YearsGeekNewsGitHub Next releases Ace, a local collaborative coding environment for humans and AI agents