Granite 4.1 LLMON previews use special-token markup to resist prompt injection
Two experimental 3B and 8B models separate instructions from data with reserved tokens, and their developers report sharply lower prompt-injection success on static benchmarks.