IBM releases Granite 4.1 previews trained to resist prompt injection

IBM released two experimental Granite 4.1 preview models, 3B and 8B, on November 12th, 2026, trained to treat instructions hidden inside documents as data rather than commands, with IBM reporting indirect prompt-injection success falling from 53-100% to 0.02-5.6% in its own tests. Basel Shbita and the IBM Granite team published the models, which are built to resist prompt injection. The release matters because indirect injection is a leading attack path against LLM applications that ingest untrusted content.