LLM agents operating in untrusted workspaces face the risk of untrusted code execution during routine tasks, yet final reports and workspace states alone can…
机构:人大
来源:arXiv 2610.40716 | AI4Papers 论文推荐平台