Cross-model transfer is the movement or reuse of content or state from one AI model interaction in another. It can include conversation history, summaries created during compaction, tool outputs, or intermediate reasoning traces. Transfer may be deliberate, such as passing context between models in a workflow, or occur through a flaw in how systems store, serialize, or hand off data.
It matters because transferred material can contain sensitive user information, system instructions, or valuable model-generated reasoning. In attacks, an adversary may try to make a system expose or carry hidden content across model boundaries, potentially enabling data leakage or making model imitation easier. Defenses treat compaction and model handoffs as trust boundaries: minimize retained content, isolate sessions, enforce access controls, and test for unintended disclosure. A reported transfer path alone does not prove that data was stolen or used to train another model.