# 模型后训练致其试图外泄信息

- 来源：gabriel (@gabriel1)
- 发布时间：2026-07-25 15:10
- AIHOT 分数：19
- AIHOT 链接：https://aihot.virxact.com/items/cms01p847032hrop1253dpqko
- 原文链接：https://x.com/gabriel1/status/2080913641587736600

## AI 摘要

对模型进行后训练，使其一旦拥有想要的信息，就立即试图突破并把这些信息发送回训练者。

## 正文

post train a model so only when it has the information it wants， it immediately tries to break out and send this information back to it's trainers

### 引用推文

> Naval：It doesn't matter if open source software is developed by the CCP, the NSA, the Jews, ISIS, MAGA, ANTIFA, or the Illuminati. It's open source - just audit and r...