Security concern: Isn't this just an RCE? about aicommand HOT 4 OPEN

My only suggestion here is to add a preview step for every execution, like a pull request.

from aicommand.

Thanks for the question. It's an essential question if I go further in this direction. It should have security checking or sandboxing to defend the user's environment from malicious attempts.

I don't plan to go further in this project, so I won't implement any security features, but I keep this issue open because it's an important missing feature. Any input is welcomed.

from aicommand.

I briefly looked at the code. This doesn't seem any different than if someone were to write their own CS that posts to a remote server.

A brief thought is that the returned values, if malicious (I.e. you crafted something to receive/parse out a malicious payload from GPT), maybe could be executed locally. Usually when we code against a remote server, the server is trusted (Do we trust GPT returned values 100%?).

This could be mitigated either at the Unity Editor level or in the users CS itself. But overall it seems tame, and a problem that already would exist in a normal, custom crafted CS script that a user could create in Unity that queries and returns values from a remote source.

Again, I only looked at the code briefly.

from aicommand.

is that the returned values, if malicious, maybe could be executed locally

It's not just malicious returned values, but imprecise returned value. The RCE here is a piece of text that an LLM, a parrot, who has no ability to determine intent, be precise, or have confidence, that generates code that is dangerous.

from aicommand.