While large language models are increasingly deployed as tool-calling agents, their safety evaluations still focus primarily on their ability to refuse harmf…
机构:Harvard University
来源:arXiv 2610.28982 | AI4Papers 论文推荐平台