Analysis
A quotation mark can belong to a document or close a field around it. IBM Granite authors use that ambiguity to explain a weakness of giving language models context through JSON: structure that a parser follows exactly can remain a learned pattern for a model.
Their November 12 article introduces LLMON, pronounced lemon, and experimental previews derived from Granite 4.1 models at the 3B and 8B sizes. Reserved special tokens mark structural boundaries. Named components and unique identifiers let instructions point to the particular data they should act on.
The boundary is the important part. Retrieved text can contain instructions an application never intended to authorize. Explicitly separating that text from the task gives the model a clearer distinction to learn; it does not by itself establish an enforceable security boundary.
The authors say indirect prompt injections are mostly ignored and attack success falls nearly to zero on some benchmarks, while general capability is maintained relative to Granite 4.1 Instruct models. The supplied excerpt does not identify those benchmarks or provide attack setups, comparison tables or independent confirmation. These remain author-reported results.
Treat the experimental previews as candidates for controlled evaluation. Who checks that each instruction is bound to the intended data? Can retrieved content change that binding in your pipeline? What control stops an unwanted action if the model follows it anyway?
Source: LLMON: An LLM-Native Markup Language, now in Granite 4.1 3B and 8B · 2026-11-12 · Primary source context · attributed account
Why it matters
Explicit context structure may help a model distinguish data from instructions, but the supplied results do not establish a security guarantee.
