if it tells me there is a backdoor, and shows me how its used, i can be pretty sure that there is one though.
if claude cant find the backdoor, chances are that claude cant find a backdoor. its still not an oracle, but that is still useful feedback for updating priors
1) The proxy is open source and much simpler than PHP runtime or a web application. Its code only takes input and generates output. It is basically a pipe. It has no other capabilities to exploit.
2) The operating system, such as BSD or Linux, supports running processes under a user and enforces limits on that user. So you just jail this proxy — since it has one purpose. Unlike your application, which may do many things and you can’t restrict it, it may become a confused deputy
What this has taught me is that we should decouple as much as possible into microservices, so that we can properly sandbox each service, and that way we have chokepoints.
Except if this code contains bugs that lead to another RCE.
> Well, this is why you should point Claude at the repo and ask it if it has any backdoors.
When do people understand that LLMs are no oracles. Just because some LLMs tell you the code is safe or has no backdoor doesn't mean it is true.