LATIDIA · Ciberseguridad

Identificar, clasificar, actuar: cumplimiento controlado por instrucción para LLM sobre datos externos

arXiv:2610.22874v1 Announce Type: cross Resumen: Los LLM sobre datos externos se enfrentan a una tensión por instrucción: un solo documento o herramienta a menudo intercala directivas legítimas con inyecciones de aviso contradictorio, y t

La noticia

arXiv:2610.22874v1 Announce Type: cross Abstract: LLMs over external data face a per-instruction tension: a single document or tool return often interleaves legitimate directives with adversarial prompt injections, and the model must follow the former while rejecting the latter. Existing defenses suppress or trust all external instructions, collapsing the safety-utility frontier and forcing a choice between over-refusal and over-compliance. We argue the missing capability is \textbf{per-instruction credit attribution} and formalize it as External-Data Instruction Classification, an identify-classify-act decision over each embedded directive. We propose \textbf{EDIC}, a decomposed reward framework for per-instruction compliance. A scene-grounded pipeline injects mixed beneficial and harmful instructions with counterfactual filtering, while a generative judge is initialized by