←back to thread

The security paradox of local LLMs

(quesma.com)
145 points jakozaur | 2 comments | 22 Oct 25 12:48 UTC | HN request time: 0.419s | source
Show context
Ekaros ◴[22 Oct 25 15:54 UTC] No.45671047[source]▶
So if you are not careful with your inputs you can get stuff injected. Shouldn't this be very clear from start? With any system you should be careful what you input to it. And consider it as possible vector.

Seems obvious to me that you should fully vet whatever goes to LLM.

replies(1): >>45671162 #
1. russfink ◴[22 Oct 25 16:01 UTC] No.45671162[source]▶
I get the impression that somehow an attacker is able to inject this prompt (maybe in front of the actual coder’s prompt) in such a way to produce actual production code. I’m waiting to hear how this can happen - cross site attacks on the developer’s browser?
replies(1): >>45671244 #
2. Ekaros ◴[22 Oct 25 16:06 UTC] No.45671244[source]▶
"Documentation, tickets, MCP server" in pictures...

With internal documentation and tickets I think you would have bigger issues... And external documentation. Well maybe there should be tooling to check that. Not expert on MCP. But vetting goes there too.