📄 中文摘要
在向大型语言模型（LLM）API发送支持票据、日志文件或用户消息时，可能会泄露电子邮件、电话号码、信用卡号码、API密钥等个人身份信息（PII）。尽管大多数开发者意识到这个问题，但由于从头构建一个数据清洗层非常麻烦，很多人选择不采取措施。为了解决这一问题，开发者创建了Airlock工具，该工具在LLM调用之前本地运行，能够有效地去除敏感信息，同时保持上下文一致性。Airlock可以替换敏感值为一致的假名，从而保护用户的隐私。
停止向您的大型语言模型发送原始个人身份信息
出处: Stop Sending Raw PII to Your LLM
发布: 2026年3月17日
📄 English Summary
Stop Sending Raw PII to Your LLM
Sending support tickets, log files, or user messages to an LLM API can lead to the unintentional leakage of personally identifiable information (PII) such as emails, phone numbers, credit card numbers, and API keys. While most developers are aware of this issue, many do not take action due to the cumbersome nature of building a sanitization layer from scratch. To address this, Airlock was developed as a local tool that runs before LLM calls. It effectively redacts PII while preserving context by replacing sensitive values with consistent pseudonyms, thereby enhancing user privacy and security.