AI Weekly Malaysia

IBM Granite published LLMON (LLM Object Notation, "lemon"), a markup language that tags each context component — instructions, retrieved data, tool outputs, requirements — with meaningful tags, reserved special-token delimiters, and unique IDs so an instruction is explicitly bound to the exact payload it acts on. They released two experimental preview models derived from Granite 4.1 base, granite-4.1-3b-LLMON-preview and granite-4.1-8b-LLMON-preview, plus a GitHub repo with a toolkit, workflow examples, and templates. The post claims indirect prompt-injection success rates drop to nearly zero on some benchmarks while general capability stays close to Granite 4.1 Instruct, but names no benchmarks and gives no numbers.

Why: If you pipe untrusted user data or tool output into a small model, this is a testable pattern rather than a product: pull the GitHub toolkit and templates and run the ID-bound instruction format against your own eval set before changing anything in production. The claim is a vendor self-report on 3B/8B preview checkpoints with unnamed benchmarks, so treat it as a hypothesis to measure, not a reason to swap models or drop existing input-separation defenses.