I'm just guessing, but seems the people who write these agent CLIs haven't found a good heuristic for allowing/disallowing/asking the user about permissions for commands, so instead of trying to sit down and actually figure it out, someone had the bright idea to let the LLM also manage that allowing/disallowing themselves. How that ever made sense, will probably forever be lost on me.

`chroot` is literally the first thing I used when I first installed a local agent, by intuition (later moved on to a container-wrapper), and now I'm reading about people who are giving these agents direct access to reply to their emails and more.

> I'm just guessing, but seems the people who write these agent CLIs haven't found a good heuristic for allowing/disallowing/asking the user about permissions for commands, so instead of trying to sit down and actually figure it out, someone had the bright idea to let the LLM also manage that allowing/disallowing themselves. How that ever made sense, will probably forever be lost on me.

I don't think there is such a good heuristic. The user wants the agent to do the right thing and not to do the wrong thing, but the capabilities needed are identical.

> `chroot` is literally the first thing I used when I first installed a local agent, by intuition (later moved on to a container-wrapper), and now I'm reading about people who are giving these agents direct access to reply to their emails and more.

That's a good, safe, and sane default for project-focused agent use, but it seems like those playing it risky are using agents for general-purpose assistance and automation. The access required to do so chafes against strict sandboxing.

Here's OpenAI's docs page on how they sandbox Codex: https://developers.openai.com/codex/security/

Here's the macOS kernel-enforced sandbox profile that gets applied to processes spawned by the LLM: https://github.com/openai/codex/blob/main/codex-rs/core/src/...

I think skepticism is healthy here, but there's no need to just guess.

That still doesn't seem ideal. Run the LLM itself in a kernel-enforced sandbox, lest it find ways to exploit vulnerabilities in its own code.

The LLM inference itself doesn't "run code" per se (it's just doing tensor math), and besides, it runs on OpenAI's servers, not your machine.

There still needs to be a harness running on your local machine to spawn the processes in their sandboxes. I consider that "part of the LLM" even if it isn't doing any inference.

If that part were running sandboxed, then it would be impossible for it to contact the OpenAI servers (to get the LLM's responses), or to spawn an unsandboxed process (for situations where the LLM requests it from the user).

That's obviously not true. You can do anything you want with a sandbox. Open a socket to the OpenAI servers and then pass that off to the sandbox and let the sandboxed process communicate over that socket. Now it can talk to OpenAI's servers but it can't open connections to any other servers or do anything else.

The startup process which sets up the original socket would have to be privileged, of course, but only for the purpose of setting up the initial connection. The running LLM harness process would not have any ability to break out of the sandbox after that.

As for spawning unsandboxed processes, that would require a much more sophisticated system whereby the harness uses an API to request permission from the user to spawn the process. We already have APIs like this for requesting extra permissions from users on Android and iOS, so it's not in-principle impossible either.

In practice I think such requests would be a security nightmare and best avoided, since essentially it would be like a prisoner asking the guard to let him out of jail and the guard just handing the prisoner the keys. That unsandboxed process could do literally anything it has permissions to do as a non-sandboxed user.

You are essentially describing the system that Codex (and, I presume, Claude Code et al.) already implements.

The devil is in the details. How much of the code running on my machine is confined to the sandbox vs how much is used in the boostrap phase? I haven't looked but I would hope it can survive some security audits.

If I'm following this it means you need to audit all code that the llm writes though as anything you run from another terminal window will be run as you with full permissions.

The thing is that on macOS at least, Codex does have the ability use an actual sandbox that I believe prevents certain write operations and network access.