The development

What happened

IBM's Granite team released experimental LLMON previews on 12 November. The AI models use reserved structural tokens and explicit instruction-to-data bindings to distinguish tasks from supplied content—not an enforced security boundary.

Key facts

  • The releases are named granite-4.1-3b-LLMON-preview and granite-4.1-8b-LLMON-preview, derived from Granite 4.1 base models. A toolkit and examples accompany them.

  • LLMON augments chat templates with addressable instruction, data and requirement spans. An execution block specifies which components belong together; reserved delimiters distinguish structure from ordinary text.

  • The authors report stronger resistance on static indirect-prompt-injection benchmarks and comparable general capability. Multi-turn adaptive attacks remain work in progress; TideReach has not reproduced the tests.

  • The team explicitly calls this an experimental preview for exploration and evaluation, not production. Instruction/data separation is learned model behavior, not a guaranteed enforcement mechanism.

The context

Why it matters

The preview tests whether context structure can make smaller AI models less susceptible to instructions hidden in documents or tool output. That is a model-design experiment, not a replacement for application permissions or independent adversarial testing.