A Guth Labs publication ↗
Back to News

IBM gives Granite an explicit instruction-data format, without promising injection immunity

IBM gives Granite an explicit instruction-data format, without promising injection immunity

IBM released experimental Granite 4.1 LLMON previews on November12, using structured prompt blocks to distinguish instructions from the data they operate on. The release includes models named 3B and 8B and a toolkit. This is an evaluation release, not a production recommendation. IBM announcement

LLMON uses reserved token delimiters and explicit references between prompt components. IBM reports improved resistance on static indirect prompt-injection benchmarks, while saying adaptive, multi-turn attacks need further work. Crucially, the separation is learned model behavior, not an enforced security guarantee. Announcement

The smaller model’s card identifies an Apache2.0 license and a November12 release. It requires the supplied LLMON chat template; ordinary chat formatting is out of distribution and can reduce quality. The card warns against using the learned privilege rule as the only defense for an agent connected to sensitive systems. Model card

Guth take: a better interface is a layer, not the whole defense

The interesting idea is to stop asking a small model to infer every boundary from a flattened conversation. If the application already knows which text is a tool result and which instruction it serves, preserving that structure gives the model clearer evidence.

But the application must get those labels right. A retrieved page accidentally promoted into an instruction block would undermine the intended design before inference begins. Explicit formatting does not replace permission checks, sandboxing, data-flow controls or human approval for consequential actions.

For local builders, the useful experiment is a controlled comparison using the same tasks and hostile inputs in both formats. Measure task completion as well as attack resistance. Test changed instructions across multiple turns and tool results that imitate trusted content. Do not advertise a complete prompt-injection solution from a favorable static benchmark.

Evidence and limits

IBM’s announcement and smaller-model card were directly read. Both are vendor evidence; no independent replication was recovered. No download or execution was performed. The names 3B/8B are product labels: the Hub’s rounded parameter display differs from the smaller card’s description, so exact tensor totals and memory requirements are not asserted. Exact announcement time is unknown.

Primary cited evidence: announcement and model card linked above, November12. Independent corroboration: none. Additional background: none. Older LLMON research exists, but this story concerns the dated preview-model release, not a claim that the research began today.

Sources and citations

Revision history

Published .