AI Agent Has Root
infernalcode.com
infernalcode.com
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.
Nobody has time to optimize for anything. They can run as fast as they can and regularly catch fire on the fly.
It's dangerous.
It's probably better to say: they're dogfooding the cheapest route to the quickfix the agent provides.
I remember a year ago we asked "Where are all the great products AI should be delivering" and now the answer is: tons and tons of AI products to consume AI, in a snake eating it's tail.
I've stopped chasing the dragon for the moment so I can push out real products.
I'm deliberately late to the game. I don't use any of the agents, harnesses, code generation, and what I see is mostly slop.
Currently, for my use case, AI is a tool for finding entrances to rabbit holes with solid references, so I can take it from there and continue digging myself.
I believe instead of diving to the deep end head-first, one needs to regulate their usage of AI, so they can get what they can get and accomplish real work with these nuggets.
I ended up with opencode and deer-flow. Deer-flow does wonders in the research and development. Opencode will produce whatever pratfalls you want.
I put out a raspberry-pi based control system replacing a expensive $$$$ PLC system with all the trimmings. Portions of it I haven't a clue how it works, but because it's not complex, I don't really care.
Prior to this, I was essentially doing the same kind of console.log/print programming, and just visually/interactively verifying the stuff I did. But there were never docs and occasionally I tried out how TDD felt, and it felt funcitonally a non-starter. I'm not a csci student.
As such, AI feels a lot more "at home" to my development than any other tools I've come across. Whenever I see an IDE that pops up method selection, etc, my entire brain breaks because like an LLM, i'm just streaming in and need to litterally see the code block or flow I'm working on. So i've hated most complex IDEs.
Interesting, our minds work completely different, then. I can generally keep most of the plan of the machinery I'm building, and working with full-blown IDEs is much more comfortable for me since it allows a birds-eye-view of the project I'm working on.
OTOH, I can't see myself attaching a LLM to my IDE and meddle with my code any time soon.
Eventually, I just use vscode to get a better picture, but i can get deer flow to pump out scaffolds, then tidy them up in open code.
It's about equal to what I can accomplish, but the current local LLMs when properly guided, do a lot better with complex dependency errors. So what I give up in my little programming bubble, I gain from not banging my head against some stupid 3rd party error or speghetti even driven madness.
They still don't get passed the level of complexity I know I can, but doing mostly greenfield or patches to open source, I spend a lot of time just restructuring.
With all that said, I'm probably in the same place as all the other AI users: not that much ROI in the bag.
Edit: I am completely in agreement with you that we need to remind people that running a lot of these harnesses/agents and LLMs gives them access to everything in your user profile. Some of works so well now, so rapidly and so painlessly that it's easy to forget about what could happen if it went wrong.
Also, we need to accept that some of the "new" problems we face are already solved well decades ago, and many people are rediscovering these solutions...
...after yelling to the very same people who were kindly pointing the right direction for being old-school, backwards and luddite minded.
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 know a lot of people are using containers for sandboxing, but given how capable the latest models have shown themselves to be for breaking out of sandboxes, I prefer the extra isolation of VMs for this.
I do all this locally - it's an interesting point to able to turn the laptop off but keep the agents running. I might consider running some of these on my homelab server just for that.
The VNC session is sized to be exactly pixel 1:1 match of my laptop screen size, so I run it as fullscreen and use the macos hot-corner functionality or multi-desktop session switch hotkeys to toggle between it and my 'real' laptop/bare metal workstation.
Having a GUI on the thing lets me leave all kinds of things running persistently in the background that might be bothersome if interrupted running on my laptop. It also has much higher average speed/reliability 24x7 internet access via my home broadband connection than wherever I might happen to be with my laptop.
The main hassle that I've found is that I have to mentally train myself to remember that a lot of the keyboard shortcuts in a boring stock xfce4 desktop environment (and other applications I can run with gnome and kde libraries installed) are quite different than MacOS, when toggling back and forth between the two.
Unless you actually need GUIs, you could just use screen/tmux or the newer versions like zellij/etc.
We’ve all been executing arbitrary code from a gorillion packages from pypi, npm, cargo, etc for well over a decade. Getting anyone to care much will be an uphill battle.
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.
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.
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.
Not sure exactly how Zed works, but wouldn't the results of the `ls` tool call be fed back into claude?
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
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
I have some bad news ... I hope you're sitting down.
How would an llm suddenly get access to your ~/.ssh folder if you didn't expressly give it access?
Like i said, its a best practice to run MCP servers, etc with their own user account and access and not with your own account.
The moment you ran your app with your own credentials you exactly granted it permission to everything you can access include ~/.ssh
It seems to be a low-effort compromise to running a VM or separate machine just for the agent, and I am still able to audit the model's work in my IDE.
People are forgetting the basic rules of operating systems. Do not run sensitive programs as a user account - especially not as a root account!
Installing "all of this was built in the last 3 months and is cutting edge" LLM stuff is very different than like, installing mariadb and then doing 'sudo mysql_secure_installation', since you mentioned database servers.
It's like comparing FedEx flying cargo airplanes in/out of Memphis in the most boring route way hundreds of times a day vs. some guy in a red bull sponsorship painted pylon racing aerobatic craft performing at an airshow in difference of risk.
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.
More importantly, publish.
Nothing that we did not already know before but this post has fooled quite a lot of HN readers.
"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.
Yes - when an entity (agent, or hacker, or whatever) gets into your account, they can do anything they want.
I thought this community understood security better than this.
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.
I'm so tired of AI voice.
Because as it damn well says - this is POSIX working exactly as designed.
I don't know why he's in here playing this up as some unexpectedly crazy, scary, dramatic music style thing (I mean... I do, it's so he can try to sell you something after scaring you...).
He started the damn process. Of course it's running as his damn user.
It would be one thing to honestly evaluate that security model and its merits (of which there are many, despite the risks). It's another to play this up as some giant surprise.
---
Constant stream of junk articles in here like
"You know when you hit that nail with the hammer you were holding? Did you consider you could have hit something else in reach instead? It's like, that hammer was like, you, man! Isn't that cray-cray and super duper scary??!?!?!"