Re:Prompt Injecting Your Way to Shell: OpenAI's Containerize...
by nilsherzig on Thursday November 14, 2024 @03:17PM
(#42140717)
They sadly do not pop a shell.
[ Parent
| on HN ]
Re:Prompt Injecting Your Way to Shell: OpenAI's Containerize...
by TZubiri on Thursday November 14, 2024 @04:23PM
(#42141382)
Yeah, they did? Inside the container, they ran ls /.
Re:Prompt Injecting Your Way to Shell: OpenAI's Containerize...
by MeetingsBrowser on Thursday November 14, 2024 @04:35PM
(#42141478)
Kind of like saying you popped a shell on replit because you ran ls. ChatGPT doesn't give direct access in the UI to the sandboxed environment, but ultimately you are still in a sandboxed environment designed to run arbitrary stuff.
The article even says so if you read all the way to the end.
Re:Prompt Injecting Your Way to Shell: OpenAI's Containerize...
by TZubiri on Thursday November 14, 2024 @04:48PM
(#42141589)
I mean the same way you ran ls, you could run /bin/bash?
Which is kind of pointless because the ls command is probably executed through a shell anyways.
So technically correct. But getting root or running a shell usually means you can run arbitrary commands, which is a check here (albeit I wouldn't say there's root access)
Re:Prompt Injecting Your Way to Shell: OpenAI's Containerize...
by MeetingsBrowser on Thursday November 14, 2024 @07:06PM
(#42142677)
There is no security implication for running /bin/bash in a service that provides you a sandbox environment to run arbitrary shell commands in.
It is like reporting that AWS lambda is susceptible to arbitrary code execution.
Re:Prompt Injecting Your Way to Shell: OpenAI's Containerize...
by TZubiri on Friday November 15, 2024 @10:34AM
(#42147814)
Yes, I too read the article.
But he did run a shell
Re:Prompt Injecting Your Way to Shell: OpenAI's Containerize...
by marcofigueroa on Thursday November 14, 2024 @04:44PM
(#42141559)
Yeah I did :)