News for hackers. Stuff that matters.

A comment on The Dark Side of LLMs Why we can not deploy the current crop of LLMs safely. Parent comment

Re:The Dark Side of LLMs Why we can not deploy the current c... by greshake on Wednesday March 22, 2023 @02:41PM (#35265364)
No, that doesn't solve it. If you run an LLM at home and give it access to APIs or your data it could still get compromised. The whole point is that it isn't the user who is doing the injection themselves.
[ Parent | on HN ]
Re:The Dark Side of LLMs Why we can not deploy the current c... by rini17 on Wednesday March 22, 2023 @05:11PM (#35267141)
But why would you supply untrusted input to AI code completion or your private AI assistant? Just treat it similarly as a shell and don't do curl|sudo stuff.
[ Parent | on HN ]
Re:The Dark Side of LLMs Why we can not deploy the current c... by wsgeorge on Wednesday March 22, 2023 @05:23PM (#35267281)
> But why would you supply untrusted input to AI code completion or your private AI assistant?

I believe that's the OP's point. Copy-pasting some text from a site into a private LLM is very easy (and fun) to do when exploring the text. When framed as "executing arbitrary code", the implications become clear. But a lot of people won't realise this until some damage has been done.

[ Parent | on HN ]