📄 中文摘要
在两个工作空间中，一个是沙盒，主要用于探索和实验，另一个是经过精心策划的作品集，面向雇主并仅限本地使用。两者之间的界限原本设计得很清晰：研究内容留在沙盒中，完成的作品单向转移到作品集中。然而，这一界限却频繁失效。发现了多个版本的 Obsidian 资料库，分别存储在不同位置，每个版本的内容略有不同。这导致脚本错误地指向了错误的根目录，作品集中的绝对路径硬编码出现在沙盒文件中，从而使两个系统之间产生了耦合，造成了上下文污染。
当沙盒泄漏：大型语言模型工作空间中的上下文污染
出处: When the Sandbox Leaks: Context Contamination Across LLM Workspaces
发布: 2026年2月18日
📄 English Summary
When the Sandbox Leaks: Context Contamination Across LLM Workspaces
The article describes the challenges faced when managing two distinct workspaces: a sandbox for exploratory research and a curated portfolio for polished outputs. Despite a clear architectural boundary meant to separate the two, this boundary repeatedly failed. The author discovered multiple copies of their Obsidian vault scattered across their machine, each with slight variations in content. This situation led to scripts targeting incorrect root directories and absolute paths from the portfolio appearing hardcoded within sandbox files, resulting in an undesirable coupling between the two systems and context contamination.