Safety in agentic systems is inherently sequential: whether an action is safe depends on prior actions, tool outputs, information disclosures, and interactio…
机构:耶鲁
来源:arXiv 2610.20960 | AI4Papers 论文推荐平台