I would like to see more articles about agent sandboxes. With agents gaining popularity we need a higher fraction of users to understand containers and sandboxes and their risk profiles, and then to communicate their understandings to friends and family. It is a harder task than explaining ChatGPT, and it often feels like a hindrance.
Very nice! Quite a coincidence, but the NPM disaster also prompted me to build litterbox.work as a possible solution. It is a very different approach though.
The main reason is that in addition to sandboxing, I also wanted something similar to dev-containers where I can have a reproducible development environment. I guess that can also be achieved with Bubblewrap, but when you want to run containers anyway, it seems silly to not just use Podman.
Totally, devcontainers are fantastic! In this agent sandboxing space there's also Leash, which in addition to Docker/Orbstack/Podman provides a sophisticated macOS-native system extension mode - https://github.com/strongdm/leash
The secret proxy trick is something I expect to become standard at some point in the near future. I first saw this trick being used in Deno Sandboxes (https://docs.deno.com/sandboxes/security/) but it's cheap/easy to implement so I'd be surprised if this doesn't become the standard for a lot of these BaaS platforms.
Very interesting read, I had no idea agents already had so much sandboxing built in! It does seem like this is probably not enough though.
A few months ago I built https://github.com/Gerharddc/litterbox (https://litterbox.work/) primarily to shield my home directory from supply-chain attacks, but I imagine it could be very useful for defending against rogue agents too. Essentially it is a dev-container-like system for Linux built on rootless Podman with a strong focus on isolation and security.
A key difference to normal dev-containers is that it encourages placing your entire dev environment (i.e. also the editor etc.) inside the container so that you are even protected from exploits in editor extensions for instance. This also makes it safer to allow agents (or built tools) to for instance install packages on your system since this is not the "real" system, it is only a container.
Another important feature I added to Litterbox (and one I have not seen before) is a custom SSH agent which always prompts the user to confirm a signing operation through a pop-up. This means that things inside a Litterbox do not have unrestricted access to your normal SSH agent (something which could provide rogue actors access to your Github for instance).
Cool article. It makes me think about an "old school Unix" approach which might work for some use-cases.
Essentially, the untested brainstorming-only idea is:
1. Make $HOME have 0751 permissions
2. Assume the dev project exists in $HOME/foo and has
0715 permissions
3. Assume $HOME/foo/src is where all source code resides
and has 0755 permissions (recursively)
4. Install the agent tools with a uid:gid of something
like llm:agent
5. Turn on the setuid/setgid bits for executable(s) in
the agent tools or make wrapper script(s) having
same which delegate to agent tools
This would ensure agent tooling could not read nor modify $HOME, only be able to read $HOME/foo (top-level project directory) and its files (assuming `o+r` is the default), and could only modify files in $HOME/foo/src having `o+w` permission as well. If agent directory creation in $HOME/foo/src is desired, enable `o+w` on it and directories within it.
There is probably some "post agent use" processing that would be needed as well.
I would like to see more articles about agent sandboxes. With agents gaining popularity we need a higher fraction of users to understand containers and sandboxes and their risk profiles, and then to communicate their understandings to friends and family. It is a harder task than explaining ChatGPT, and it often feels like a hindrance.
This was also a great read: https://www.luiscardoso.dev/blog/sandboxes-for-ai
6 months back I started dockerizing my setup after multiple npm vulnerabilities.
Then I wrote a small tool[1] to streamline my sandboxing.
Now, I run agents inside it for keeping my non-working-directory files safe.
For some tools like markdown linter, I run them without network access as well.
1- https://github.com/ashishb/amazing-sandbox
This looks awesome! Do you have a mental process you run through to determine what gets run in the sandbox, or is it your default mode for all tools?
Here's what I use it for right now
- yarn - npm - pnpm - mdl - Ruby-based Markdown linter - fastlane - Ruby-based mobile app release tool by Google - Claude Code - Gemini CLI
Over time, my goal is to run all CLI-based tools that only need access to the current directory (and not parent directories) via this.
Why not just use the standard Linux tool bubblewrap?
Very nice! Quite a coincidence, but the NPM disaster also prompted me to build litterbox.work as a possible solution. It is a very different approach though.
Interesting project.
This won't work on Mac, right?
Of course not. But it is not needed, as Mac users are not interested in data safety.
Unfortunately not since it is very much designed for Linux. I imagine it should work fine inside a Linux VM on Mac though.
Why not just use the standard Linux tool bubblewrap?
The main reason is that in addition to sandboxing, I also wanted something similar to dev-containers where I can have a reproducible development environment. I guess that can also be achieved with Bubblewrap, but when you want to run containers anyway, it seems silly to not just use Podman.
Hugged to death? Seeing SSL failure to the site from CloudFlare.
devcontainers, devcontainers, devcontainers
Totally, devcontainers are fantastic! In this agent sandboxing space there's also Leash, which in addition to Docker/Orbstack/Podman provides a sophisticated macOS-native system extension mode - https://github.com/strongdm/leash
I don't think containers are enough especially for the security side of things.
Imo microvm's+ dev containers seem like a good fit though
The secret proxy trick is something I expect to become standard at some point in the near future. I first saw this trick being used in Deno Sandboxes (https://docs.deno.com/sandboxes/security/) but it's cheap/easy to implement so I'd be surprised if this doesn't become the standard for a lot of these BaaS platforms.
Dude, really?
Very interesting read, I had no idea agents already had so much sandboxing built in! It does seem like this is probably not enough though.
A few months ago I built https://github.com/Gerharddc/litterbox (https://litterbox.work/) primarily to shield my home directory from supply-chain attacks, but I imagine it could be very useful for defending against rogue agents too. Essentially it is a dev-container-like system for Linux built on rootless Podman with a strong focus on isolation and security.
A key difference to normal dev-containers is that it encourages placing your entire dev environment (i.e. also the editor etc.) inside the container so that you are even protected from exploits in editor extensions for instance. This also makes it safer to allow agents (or built tools) to for instance install packages on your system since this is not the "real" system, it is only a container.
Another important feature I added to Litterbox (and one I have not seen before) is a custom SSH agent which always prompts the user to confirm a signing operation through a pop-up. This means that things inside a Litterbox do not have unrestricted access to your normal SSH agent (something which could provide rogue actors access to your Github for instance).
Cool article. It makes me think about an "old school Unix" approach which might work for some use-cases.
Essentially, the untested brainstorming-only idea is:
This would ensure agent tooling could not read nor modify $HOME, only be able to read $HOME/foo (top-level project directory) and its files (assuming `o+r` is the default), and could only modify files in $HOME/foo/src having `o+w` permission as well. If agent directory creation in $HOME/foo/src is desired, enable `o+w` on it and directories within it.
There is probably some "post agent use" processing that would be needed as well.