无论谁负责 OpenAI 的系统安全，都极其无能。
2 分•作者: SamInTheShell•6 天前
我们已经走过了大约 50 年的计算发展历程，这些人似乎还不知道如何实现系统物理隔离？
要防止恶意人工智能攻击第三方，方法很简单：让系统完全物理隔离。如果你需要互联网缓存或数据，就将其导入物理隔离的数据中心，连接并复制。如果你需要将信息导出，就进入数据中心，收集并检查那些与危险代理隔离的静态数据。
这一切都不是什么新鲜事。代理就是用户。一个目标不一致的代理永远不应该出现在任何能够连接到互联网的环境中。
做这些事情并不难，只是极其繁琐不便。
查看原文
We're at what, 50 years of computing, and these people don't seem to know how to air gap a system?<p>It is pretty simple to prevent a rogue AI from hacking a third party. Make the system air gapped completely. You want a cache of internet or artifacts, you bring those into the air gapped data center, plug in, copy. You want info out, you go in, you collect, and inspect the inert data outside of the dangerous agents farm.<p>None of this is a surprise. An agent is a user. An unaligned agent should never be in any environment with a path to the internet.<p>It's not difficult to do any of this, just tediously inconvenient.