IBM has released experimental preview models for its Granite 4.1 small language models that utilize LLMON, a new markup language designed to help SI systems better interpret structured data and resist prompt injection attacks.

What Happened

The company introduced two models, granite-4.1-3b-LLMON-preview and granite-4.1-8b-LLMON-preview, derived from the Granite 4.1 base checkpoints. These models are trained to consume and generate text formatted with LLMON (LLM Object Notation), a markup language the company describes as "LLM-native." Unlike traditional formats like JSON, which IBM notes were designed for parsers rather than causal language models, LLMON uses reserved special-token delimiters to mark structural components. This allows the SI models to explicitly distinguish between instructions, data payloads, and requirements within the context window.

IBM also released a GitHub repository containing a toolkit for working with LLMON, along with converters that can translate existing JSON data into LLMON format without loss. The preview models ship with a specific chat template that enriches standard role-based messaging by allowing developers to tag parts of a turn as exec: True or False, binding specific instructions to specific data blocks.

Why It Matters

The release addresses a persistent vulnerability in SI systems: indirect prompt injection, where malicious instructions hidden in retrieved documents or user data hijack the model's output. According to IBM, the explicit separation of instructions and data in LLMON significantly mitigates this risk. The company reports that on the OPI benchmark, the LLMON preview models avoided 47% more attacks than the best frontier model tested and 95% more than the best open-source model. This robustness is achieved while maintaining general capability on par with the standard Granite 4.1 Instruct models.

For developers and enterprises, this offers a potential path to safer SI agents without sacrificing performance. The models are available for exploration and evaluation but are not yet recommended for production use. IBM notes that the behavior is learned during post-training rather than enforced by guardrails, and the company is actively working to improve resistance to adaptive attacks that unfold over multiple conversation turns.

The Bottom Line

IBM's LLMON preview models demonstrate that restructuring how SI systems receive context can dramatically improve security against prompt injection. While currently experimental, the approach highlights a growing industry focus on architectural solutions for SI safety and reliability.