AI Agent Has Root

This article smells AI generated, which is _very_ funny given the argument being made.

Either way, while this is true in the absolute, this is the value of building good MCP servers: they should expose only exactly the surface you expect your agent to need, and adding functionality should be carefully considered. The best MCP servers I use day to day (Cloudflare sticks out) do a really good job of exposing only what an agent might actually want to do on my behalf, rather than just all and sundry. Unfortunately the Chrome Dev tools MCP is less discriminating and is only as secure as a browser sandbox with full JS access (not fatal but not as strong as a well scoped REST API).

All of this is sidestepped somewhat by using good isolation primitives - I'm running a Hermes agent as of recently on a DigitalOcean VM that only accepts connections from my devices over tailscale, and it has all its own credentials so I can revoke them easily should they be used maliciously. Giving an agent root on a box is not _necessarily_ a huge deal, you just have to make sure that box has nothing valuable on it.

I kinda feel like we're rediscovering "Cattle, not Pets" when it comes to the environments we run our agents: give it root, sure, but a root that is almost meaningless outside of the functionality you granted it.

This is why I run opencode and similar things in a dedicated KVM virtual machine that lives on a system under my desk that doesn't have access to my user account data, documents folder, photos/video, ~/.ssh/, other API keys, ~/.anything-else/, you name it.

I think it's absolutely wild that there are people out there running cutting-edge LLMs and agent harnesses and tools on the same hardware and same user/disk/session environment that contains like, PDFs of their paystubs, their 401k records, their tax returns for previous years, contracts/real estate details, whatever other personal things you keep in your Documents folder as an educated modern professional with obligations and debts and assets.

Sometimes this VM gets duplicated for specific projects and then various dependencies for testing installed in it that are specific to what the needs are.

As an additional advantage it means I can leave it running in the background doing things when I want to shut my laptop, then resume talking to it later.

(edit, for everyone who hasn't seen it yet, take a look at the "grok uploads your entire code base" category of problem: https://www.google.com/search?client=firefox-b-d&q=grok+uplo... )

I once opened Zed, howdy there was an agent window on the side, I log into Claude. I asked it to list what it could see and work with... And there my (priv) ssh keys come scrolling by. I don't know what I was expecting really, it's very logical, but also very in your face. They didn't leave my computer I guess, I just showed me the output of some `ls` commands. But still.

So now I have CC in a container, mounting it's own credentials/memory folder per project and only mounting one repo at a time (script in the repo itself). I feel a bit better about it now. It can still access my networks of course.

So this looks like Claude is discussing the idea that when you run untrusted code on your computer, it can do arbitrary things. If there is malicious code, maybe it'll read your files and steal your credentials.

Here's a sample: "That’s not hypothetical. That’s POSIX working exactly as designed. None of it requires exploiting anything. It’s your user account doing normal user account things."

I would recommend clicking TFA if you haven't heard of a supply chain attack, or if you just like getting the honest load-bearing facts that are worth mentioning.

I've created boxy [1] to sandbox the agents via Landlock and gh-proxy [2] to not share GH PATs that might leak over the internet.

I have a far better setup on my Kubernetes cluster [3], but these are good building block (IMHO) to start preventing these kind of issues.

I also "recklessly" run `claude` / `codex` as root for certain things - but that happens on a completely separate machine that is meant to be pruned afterwards, and it's what unlocks the kernel development feedback loop that is needed to port a device (such as the Daylight DC-1 / Surface Pro X) to mainline Linux.

[1]: https://github.com/denysvitali/boxy

[2]: https://github.com/denysvitali/gh-proxy

[3]: https://blog.denv.it/posts/im-happy-engineer-now/

[4]: https://x.com/DenysVitali/status/2091238391710888416

Author of the post, and dev of mcp-box here. Wrote this after realizing every MCP server on my machine had the same access to ~/.ssh that I do, and nothing in the installation messages posting to stdout mentions it. I think prompt injection is the vector and the permissions model is the red carpet giving a warm welcome. Interested in where that's wrong.

Expecting an agent harness to respect security boundaries is definitely the wrong thing.

We need to have easy to use auditable sandboxing of agents and their harnesses, because it's quite clear certain model providers are going to try to lock us in at the harness level, and it's those harnesses that seem most likely to go nuts.

Always put things in a container and ensure its safe before using something like AI! I put my opencode agent (with full permissions on an up-to-date debian VM with a DMZ virtual network configuration). If it figures out how to get out, then I'll have fun documenting that at least!

That's exactly why I'm migrating all my secrets to infisical (of course I guess there are many other provider as well), securing my ssh keys with Ubikey and in general "refactor" my security concept. Still much today and I doubt that I would stand a chance against a frontier model trying to hack me but I don't see an easier alternative.

After the grok cli fisaco (uploading the full repo) I moved my different projects to seperate users, seperate sandbox around each harness (nono). Vm is of course even better but resources and easy of use start to be an issue

I use aicommander.dev for managing all hardware via agents

I think it’s safest to never run AI in your own machine. I only run it in Linux VM’s I created on exe.dev.

Other advantages: the AI doesn’t stop working when I close the laptop lid. I also prefer running the AI in a browser tab to a terminal.

step 1) create a Dockerfile with claude-code / pi.dev + acp_agents + whatever dev env you need. step 2) use microsandbox to create a fs image from that Dockerfile image. ( All secrets you need inside the vm live only on the host and are injected in the tcp stream on egress) step 3) microsandbox by default speak ssh without you needing to have a ssh server in the vm. You can proxy forward everything through it if you need. In my case i use zed and connect into the vm. Bonus points all zed "agent" stuff now runs inside the vm. E.g zed threads via acp calls and all agent stuff live only in the vm.

how is this any different to naively running a webserver or database server or, in fact, anything as yourself?

People are forgetting the basic rules of operating systems. Do not run sensitive programs as a user account - especially not as a root account!

Did you even review this? Holy moly. "I'm just trying to generate content that’s informative and useful, to someone, anyone". Come on now, you're joking right?

"or send a paranoid infoSEC analyst into a full lockdown and round the clock obsessive kernel hardening sessions."

Please look inward. You're saying you want to 'generate' useful content. You are doing harm. This is vague, confusing, and filled with incorrect or inaccurate information. There is of value here. You are completely unclear of what tools are being used as examples.

I've typed and deleted a sentence here 5 times in a row before realizing im putting more effort into explaining the slop, than seems to have been put into reviewing the content of the article.

Edit: I cant help myself. cue the dramatic music Who are you writing this for? What person exists that will come away from this with actual value? I'm willing to bet that a user who doesn't already consider this stuff, will also think you mean Claude Web when you say "claude" without explaining which damn tool you're talking about.

I implore you to stop, close your eyes, take everything you know about users, permissions, and start thinking harder about the implications.

a discussion from a little while ago where many of the comments discuss sandboxing strategies: https://news.ycombinator.com/item?id=49239751

> Install things via pip, npm, cargo, wherever you have write access

More importantly, publish.

This entire blog post is AI generated with the most obvious AI patterns and this post was totally botted to the top of HN.

Nothing that we did not already know before but this post has fooled quite a lot of HN readers.

Wouldn’t it be safer to set up another user entirely for AI agents?

A fairly comprehensive list of ways you should be sandboxing your agents: https://pleasedonotescape.com/

If you like a real focus on developer experience and want lots of bells and whistles that a lot of thought has gone in to, I wrote byre, which I think is very good: https://github.com/pjlsergeant/byre

you can run the agent in a different environment

I run these things in VMs, and as a user in the VM, and give them access to just the tree I want to work on via mounting. I use Parallels on macOS for this but any VM will do.

That way my actual primary machine is at least strongly isolated. I also run them on VMs on my home Proxmox box.

Zed, as an IDE, can connect to a VM via ssh and both edit and run things like Claude or its built-in agentic harness on your code with the stuff running in the VM.

That way if an AI goes Skynet all over the machine, it's at least isolated. There's malicious injection attacks but there's also edge case failures I've seen where the AI decides it "needs" to do something asinine.

>That’s not hypothetical. That’s POSIX working exactly as designed.

I'm so tired of AI voice.