IBM Open-Sources Small Models Trained to Resist Prompt Injection
IBM released two small open-source models (LLMON preview) trained to ignore hidden malicious instructions inside documents, cutting indirect prompt injection success rates to nearly zero on two public benchmarks. The models are especially useful for RAG and agent pipelines that read webpages or user-uploaded files.
2026-11-14 ~ 2026-11-14 · 2 related posts
- IBM trains small open models to ignore hidden injection attacks, cutting success to near zero — mawcopolow · 2026-11-14
- IBM ships two small open models trained to ignore prompt injections in documents — MajesticPuzzler · 2026-11-14