Preventing cross-entity contamination in AI outputs