Modern AI agents routinely cross trust boundaries: they ingest untrusted content, combine it with privileged instructions, persist intermediate beliefs in lo…
机构:清华
来源:arXiv 2609.29647 | AI4Papers 论文推荐平台