IBM previews Granite 4.1 models built to resist prompt injection with LLMON
4.7/10Significance
IBM published on November 12, 2026 two experimental Granite 4.1 previews, 3B and 8B, trained on LLMON, an LLM-native markup language. IBM says they avoid 47% more attacks than the best frontier model it tested.
What changed
IBM says LLMON on its OPI benchmark avoids 47% more attacks than the best frontier model tested and 95% more than the best open-source model.
Why it matters
Prompt injection is a main obstacle for AI agents, and this is an open attempt to handle it in the model itself.
What remains uncertain
The results are IBM’s own, on one benchmark. IBM calls the release an experimental preview, not ready for production.
What to watch
Watch for independent tests of the preview models.
Source
Why this ranks here
A notable open security approach, but self-reported and experimental.
- Impact4
- Reach4
- Novelty6
- Commitment5
- Coverage quality5
- Future potential6