Key Defense Pillars三大防御防线

1. Separation of Model and Execution1. 推理与执行权严格分离

The AI model has strictly zero shell, filesystem, or API token access. All user data modifications are performed by deterministic Worker TypeScript code after verifying JWT/session tokens.

大语言模型自身绝无 Shell 命令行、文件系统或 API 凭据访问权。所有数据读写均由确定性 TypeScript 代码在严格鉴权后执行。

2. Untrusted Prompt Sanitization2. 不可信输入过滤与防注入

All user queries, retrieved memories, and external inputs are treated as untrusted data. Prompt injection patterns and system override keywords are blocked before reaching inference.

所有用户输入、检索出的记忆碎片与外部内容均视为不可信数据，在进入大模型前均经过严格的关键词与注入检测阻断。