Separating Commands from Data with Exclusive Codes: IBM Tests a Structuring Language to Protect Granite Models from Indirect Breach
Listen to this article
Read by Anchor
Language models build their responses on the context available to them, but in complex work environments that context becomes a crowded space where system instructions mingle with conversation history, retrieved data, and external tool responses. The core problem is that the structure of this data is usually passed in formats such as JSON, which were designed for traditional software parsers rather than for attention mechanisms in generative language models; the lexical tokenizer treats quotation marks and programming brackets as generic textual characters, which confuses small models in particular and opens the door to indirect command injection attacks embedded in user documents.
To address this structural flaw, IBM Granite’s research team introduced a custom description language for language models called LLMON, alongside two prototype previews derived from the Granite 4.1 family with sizes of 3 billion and 8 billion parameters. The idea is to embed six exclusive delimiters within the lexical tokenizer’s vocabulary itself to precisely delineate text boundaries and component semantics without ambiguity.Input components here become independent blocks carrying binary-character identifiers, linked to a specific execution block at the end of the context that precisely defines the command to be executed and the data it operates on.
This architectural separation prevents data texts from directing the model; instructions embedded in a retrieved document or tool output are treated not as executable commands but as plain readable material. Evaluation results on the OBI benchmark for indirect command-injection resistance show that these experimental models resisted attacks almost completely, outperforming the best closed-source advanced models by 47 percent and the best open-source models by 95 percent, while maintaining overall inference efficiency.
The new behaviour was not added as an extra protection layer on top of an existing model; it started from the core inspection points of Granite 4.1 by re-framing training data and enriching it with obfuscation instructions and data deliberately left unlinked to the execution block so that the model learns to ignore them automatically.This was followed by a five-stage reinforcement training sequence using a GROP algorithm that included reasoning, knowledge calibration, and computation, culminating in a direct preference-optimization stage based on a thousand pairs of real responses.The architecture also allows software teams to create custom description vocabularies and train compatible LoRA adapters, as demonstrated on the Science World benchmark for scientific task simulation.
For institutions and technical teams in the Gulf, Egypt, and the Levant, this trend has tangible operational relevance; many banks and government agencies in Riyadh, Dubai, and Cairo now adopt small and medium open models for deployment on local servers to comply with digital sovereignty, financial, and governmental data protection. However, relying on JSON processing and merging it with unstructured text exposes automation agents to the risk of compromise or collapse while extracting fields.Practical hardening against vulnerabilities in smart agents does not start with additional external filters that slow response time, but stems from the semantic structure the model learns within its own context.
Although the research team emphasized that the two models are experimental previews intended for testing and evaluation rather than for commercial production, providing programmable JSON conversion tools and integrating them with the Maliya library gives developers the opportunity to examine attention behaviour and experiment with separating data from guidance. This step alerts local smart-app developers to the need to review how they pass and prepare context before incurring operational costs in environments that demand strict protection.