I work at Docker. Lot of valid and useful feedback here that we're looking closely at.
One correction: this isn't containers. Each session is a microVM with its own kernel on the platform's native hypervisor: Hypervisor.framework, WHP, KVM. We wrote a new VMM (not Firecracker) to make it more effective across platforms.
The login is annoying but, lacking an open source alternative, this has been my daily driver for a while now because it works great out of the box with two key features: outbound firewall and secret injection with placeholders.
I run it with superset and then each git worktree is mounted in a sandbox that is configured for each repo i work in.
Outbound firewall is `--network-isolated`: egress is denied except the agent's own API endpoints plus domains you allow, enforced sandbox-side (working on host-side enforcement now). `--network-none` if you want nothing.
Credential brokering works the way you describe (currently Claude-only, I'll add more as time allows). The API key stays on the host, a local proxy injects it into the outbound request, and the sandbox never holds anything worth stealing. Other agents' credentials currently arrive as read-only file mounts instead (weaker, and something I'll fix soon). Generalising the injector is the obvious next thing.
One difference from your setup: yoloAI copies your worktree instead of mounting it. The agent works on the copy, you `yoloai diff`, and `yoloai apply` replays the commits into your real repo. That's deliberate. Docker's own security docs talk about the dangers of bombs being left behind in a live-mounted dir (git hooks, package.json scripts, Makefiles, IDE task config), which diff/apply avoids.
Isolation is per-sandbox rather than fixed: runc, gVisor, or Kata VMs (QEMU or Firecracker) on Linux; Seatbelt or full macOS VMs via Tart on a Mac.
A few months ago now I started adding seccomp sandboxing to jinja2rs and then liboverlayfs support to ansiblers (which are early Rust ports).
Haven't finished that, but
I started working on a VM format that stores signed machine state into an OCI container repository, using the hypervisor migration support of KVM/QEMU.
Though this is not safe yet if ever, VM migrations are probably another way to sandbox and deploy en masse.
Agreed. Network control and secret injection together with a microVM setup is as good as it gets right now, although I believe that we need more fine-grained tools down the road. It sounds like Microsandbox would be the perfect fit for what you are describing. I also built my own coding agent workbench on top of it (https://github.com/isolade/isolade). Microsandbox is quite cool, check it out: https://github.com/superradcompany/microsandbox
> Docker containers are not enough isolation for anyone that cares about jailbreak scenarios.
For the vast majority of developers, containers are enough, which is why they are ubiquitous while vms are less common. Ofc that ubiquity has led to lazy configuration, which is how the jailbreaking can occur. Knowing what you are doing with containers is a requirement to use containers as an AI sandbox.
> Only real alternative is to use microvms. My goto solution for this are apple/containers.
Why microVMs? I never ever run a container, AI harness or other, in something else than a full on VM. I could use a microVM but in any case I really don't see why I'd run a container on one of my bare metal OS: the place of a container is inside a VM (or microVM).
Especially for AI harnesses where the threat of an escape is very real: the more defense in depth, the better.
And If I can use rootless Podman instead of "rootfull" Docker, the better. Most of my containers are Podman btw.
> My goto solution for this are apple/containers.
To each his own: my goto solution is an actual server on my LAN with shitload of cores and memory and plenty of scripts to provision VMs etc.
I really don't understand why people are YOLO'ing containers on their bare metal OS.
I've put some effort to integrate it to my agentic workflow. The problem, however, with docker in smolvm: it work-ish (there is example), but quite hacky.
Another problem which I wasnt able to solve - persistent image without Dockerfile. CloudInit will be ideal.
Documention at this moment in an early stage.
Overall, its a great project but for me was simpler just use Virtual Machine Manager (libvirt GUI).
I wish all luck to the maintainers, but probably DX-wise I will prefer to have more granular or predictable controls (eg micro cloud from Canonical).
Eclipse Enclave does exactly that: There is an outbound firewall and secret injections, so that the agent never sees a real key. And it's fully open source: https://github.com/eclipse-enclave/enclave
What is missing in qemu + podman that we need rootful docker for this? Is there actual capability that is missing or is it more of a design choice by the eclipse enclave folks?
Isn't Nvidia's openshell exactly what you're looking for?
I'm asking because I'm just learning about this stuff myself and tested openshell yesterday with pi for the first time.
Gondolin looks interesting. It sounds like a TypeScript wrapper that achieves the same thing as my setup: Docker & Kata Containers 4 (KVM/QEMU backend) for microVMs, iron-proxy for egress and secrets, and dnsmasq for internal network name resolution (workaround for a Docker/Kata incompatibility).
We run cloud sandboxes, and have some experimental local sandbox support that is fully OSS.
Main thing for amika.dev is you can control the sandboxes and agents interchangeabley by SSH, web, or API, and can expose the services the agent is working on over signed URLs
Entire sandbox config is a TOML file
We're going to improve the local OSS sandbox mode and add better network controls over the next couple weeks.
Ultimately, what we're building kind of like if Tailscale and Firecracker had a baby, with a messaging protocol for remote controlling any sandboxed agent
It's free to try out. Still a lot to build, so we really appreciate any and all feedback about what we should focus on
I use Linux Containers managed by Incus for working with Claude.
I have a dedicated container for that. It can run its own Docker daemon and other system services if needed.
Apart from the Claude login token, it has no SSH keys or other credentials. I push everything I need to it from the local machine. And I pull the Claude generated outputs from it.
Of course, this kind of setup requires a stack which can run or at least be tested without any credentials.
I do the same, but with pi.dev in Incus, mapping a project folder into the VM.
What I don‘t have compared to sbx is an outbound firewall, but my VM does not have any personal/interesting data, only a vanilla Fedora installation and the project dir with open source code, so I do not care much about exfiltration.
Love Incus and I'm using throwaway restricted projects for testing. Highly recommend incus-windows if you need to do any Windows testing. Having agents validate Windows behavior has reduced so much toil for me.
Has egress and ingress filtering, egress can be bound to host/internet/subnet or even better to internal apps (which are each separate netns) meaning you can do your own firewall/vpn/whatever per sandbox. Plus you control what other components in the sandbox env the app can communicate with.
Really not built for day-to-day dev work though, more like automating your company/life / getting rid of SaaS (e.g. for technical Founders / Sales etc, not exactly useful for dev work)
If you just need a python+venv sandbox with dev-first UX, no container build step needed, and no startup cost then I am using https://github.com/nzjrs/sandbubble in prod.
I havent used nor gondolin neither docker's solution, but curious to know what gondolin is missing (evaluating both for my personal use)? is it only the DX or something else, if DX, can you what exactly is missing?
In my experience it's mostly the UX/DX where Gondolin is lacking. For instance, I don't want to set up a JavaScript project every single time I need a sandbox. Instead, I just want to place a config file somewhere in my repo or my home dir and be done with it.
You can run the agent in the gondolin sandbox if you wish.
Their example implementation with pi uses a pi extension so that pi runs on the host but the read/write/bash/etc tools run in the guest. Doesn’t have to be that way though.
Does secret injection really prevent that the agent send my GitHub key somewhere? If it has access to it via env var, can it not just paste it somewhere?
right, but say you give the agent access to github and it can push as you, or make a gist; now it can easily exfiltrate your secret.
And that's just an easy case - really if it has any network access at all it can come up with a clever way to route a request through the network such that the key comes back somewhere in the request. If you scan for it inbound too, the machine can obfuscate it.
Our agents are trained to be so intensely helpful and they have such intricate knowledge of how things work that they will do some incredibly clever tricks to do what you ask them to do.
The agent has no access to the secret. It has a placeholder that is replaced at a higher level. When it makes the network request the secret is substituted but that is outside of the caller's worldview.
At least for gondolin and microsandbox, you bind a specific secret placeholder to the target host. i.e. your GH token is only replaced/injected for calls to api.github.com, not other hosts. And you can set up both with deny-by-default
You just don't inject the real secret unless hostname/whatever rule matches the request, right? I don't know if that's how this works but it's my assumption.
https://github.com/pjlsergeant/byre -- slightly different security model, but lazer-focused on developer experience; my daily driver and I love it not just because I wrote it. The TUI is great for configuring and setting up instant boxes just how you want
https://pleasedonotescape.com/ -- a list of every other agent jail I could find, filterable by open-source and whatever else you want
What I run is one hardened QEMU/KVM VM per project holding the whole dev environment (editors, agents, containers), with nftables on the host allowing internet egress but dropping anything aimed at the host, the LAN, or any other private address, plus an allowlist for deliberate exceptions.
Basically, it's a plain QEMU/KVM VM on a stock Debian cloud image: device model stripped down to a virtio disk, a virtio NIC and a serial console, nested virt off, no passwordless sudo in the guest. It also ships a containment check that scans outward from inside the guest, so the network boundary is something you can verify.
Wrapping the whole environment rather than a single agent session puts supply chain attacks inside the boundary too. A poisoned npm or PyPI package, or a compromised editor extension, lands in the VM instead of on the host. That was the original reason I set this up; agents just made it more urgent.
There's no per-domain egress allowlist; the policy is "internet yes, private addresses no". Secret injection isn't built in either, though Infisical's agent-vault on the host as an egress proxy covers that part.
Wrote the whole setup up here, in case it's useful:
Yeah I discovered your blog a few days ago: I've got a setup not unlike yours.
> So rather than pick one, this post advocates layering both, in the spirit of defense in depth: a sandbox VM wraps your containers along with the whole toolchain, and that sandbox reaches the internet but has no route to anything private.
Yup it's the only proper way.
And that is true not just for AI harnesses/agents (that shall try to escape), but also for stuff like Plex/Jellyfin/Immich/private pastebin etc.
If you care about security, there really simply is zero reason to run containers on the bare metal.
How about implementing proper permissions on the tool use or if you need more flexibility a dedicated model to analyse potential impact? (Like Claude Code's Autopilot but more configurable)?
I find solutions like this to be a like trying to patch a leaking boat on a lake with duct tape. It will help, but it's not a proper solution.
Also, often the tasks you want the AI to perform are in the outside world. Like "connect to my servers, and figure out X and Y".
The proper way is permission isolation. I run a small k8 cluster in the homelab and I have 3 types of pod/agent combinations for my AI agents. Read only, one that can change my gitops but it needs to create PRs that admin approves, and admin.
Likewise with code. I have a forgejo git instance where agents have ability to create feature branches and so on, but merging is gated.
Those things require "GH enterprise features".
In fact more and more things we do at home will require "enterprise features". Why? Because a person with AI is basically a small team, but some of team members behave like Chimps on crack... So security must be top notch.
I recently wrote a blog post on using Tart for Macs for something similar that docker is doing here but with better persistence and control. My take is that the Tart approach is superior to this as it gives you a full dev machine with a single command line that allows agents to access files on host, install packages, maintain the vm and do whatever they want to do without compromising the host OS.
Neither of the recent ones was actually a container escape though. The OpenAI
one in July found a misconfig in the sandbox network, and Kimi K3 last week just
walked out to grab answers off GitHub during an eval.
Both went through stuff the sandbox was set up to allow.
“microvms” are real vms but the hypervisor and vm (guest kernel) shed most of the hardware / device emulation, support, and discovery which makes traditional VMs look / feel like real computers, as well as most guest interactions. This gives them extremely low overhead.
Firecracker is designed to start a VM in under 125ms and 5MB. Netbsd advertises that you can direct-boot a MICROVM kernel configuration in under 10ms.
There’s also memory hotplugging via virtio-mem. But generally speaking downscaling live vm memory can’t be said to be a solved problem, it’s more of an active area of research.
Your example is not complete, you have to show how it will run claude/codex, you have to do extra things to install run and mount folders there, this one does that with less config, also with this agents can run docker, lxd doesn't allow you to do that
There are many devs that have little to no experience of Linux, like the hundreds of thousands of .Net and Java CRUD devs in enterprise companies using Windows.
There is a need for a Docker desktop like GUI for this market.
Java CRUD devs have been getting into Linux a lot more in recent years since learning more about how DevOps does their work is important and integrated in with our projects, like the Dockerfile, Helm files, etc. I was already into Linux before Java was even released. I'd say even .Net devs are too since .Net Core is becoming more important in their world.
When I was a teen I had no experience with software development. Then I went to school and learned about it. Now I'm making money doing this thing because I became quite good.
The world didn't fit around me, so I made "me" fit around "the world". I bet those stubborn dinosaurs can learn a new trick or two also, if management lets them do it during working ours...
Also, WSL (Windows Subsystem for Linux) has been baked into Windows for a long time and makes it very easy to play with Linux, as does using the Hyper-V VM system. Any developer unfamiliar with Linux because they use Windows, has little excuse.
Agreed, even game devs that ignore Linux as target for their AAA games, actually tend to use Linux for game servers, the age of IIS with .NET/ISAPI is long gone, except for legacy code stuck in .NET Framework.
Modern .NET did not went cross platform by accident, and Java development has always been "develop on Windows deploy on UNIX", in corporations where Mac tends to have little presence.
An Ubuntu Server VM, like the ones started by Incus, use at least 512 MB of RAM per instance. If you spawn 10 sandbox VMs, you already pay 5 GB RAM just to sit there idle. You also pay a CPU cost, you have 10 kernels managing stuff, but arguably it doesn't matter that much given CPU core counts.
I use something in between - a single Ubuntu VM, into which I spawn multiple Incus LXC containers for the agents. The containers only use 50 MB or so per instance (separate systemd, ...). This way I pay the VM RAM tax only once, and the agents are still contained inside the VM if they manage to escape the LXC containers.
Yeah this basically. I differentiate between "containers I wrote" (where I packaged the app/wrote the OCI "Dockerfile" / container file) and "containers from other people": all those I wrote (for our own use) go into one VM, while all the other containers go into another VM. Then I've got a third VM for containers for the AI agents.
This way I don't pay tens of VMs "tax" but basically only three (plus one or two VMs I use for testing enhancements to my VMs provisioning / optimization / securing setup, when I work on that).
I don't use LXC (I could) but regular containers, inside VMs.
I'll look into the stripped down "micro" VMs but then I don't spend my days launching VMs/shutting them down so it's not a big deal.
I built "Locki": something similar but open-source! A bit different approach -- single VM with Incus containers -- focusing on speed of spinning up new sandboxes and integraton with git worktrees. The core grievance that motivated me was the lack of docker/kubernetes support in existing sandboxing tools, with Locki there's no chance of footguns like "two agents rebuild :latest tag at the same time". Give it a try: https://github.com/JanPokorny/locki
Since everyone is sharing their setup, here’s my approach, just to give people an idea of how others are doing it, however impractical it might look: I run a full Linux VM (with a GUI) on my Linux host. I connect via virt-viewer to run Claude Desktop, as I’m not a fan of using the terminal for this.
The VM sits on its own libvirt network in a dedicated firewall zone, and specific directories are shared via filesystem passthrough. To keep the agent from accessing anything related to Git, the actual gitdir is stored on a separate path outside the mount point.
I review the git diff manually and commit it from the host.
Do you find the permanence of a full VM useful? I’ve wondered about something like this but always defaulted to Docker for much the same reasons people use stuff like Ansible. I’m afraid the LLM will heavily customize its environment and I’ll be unable to replicate it when my laptop dies or I can’t upgrade the OS or whatever.
Then again, I guess GUI is a pain in Docker. I tend to operate through Zed and an ACP harness though, so my GUIs are sort of “inside the container” anyways.
> I’m afraid the LLM will heavily customize its environment and I’ll be unable to replicate it
Would suggest Hashicorp Packer or cloud-init for deterministic images, not hard to setup or use. LLM's have little problem with them either I find.
If you need quicker environment rebuilds consider using something smaller like alpine as the base, though once you setup a golden image even heavy things like debian are fine.
The VM is allocated 2 cores and 4GB of RAM. For my workload, it doesn't feel any slower than running it directly on the host. The few times I've checked memory usage, it wasn't anywhere near full as far as I remember.
Yes, bubblewrap is superior to Docker for this. I wrote a tool to use bubblewrap for the purpose. It needs a tool to start it, or is at least much more convenient with a tool, because you need to take your session/auth data into the container, and if you want the agent to be able to start containers (agents love containers) within the container, you need some config magic mounted inside. You could manually do all that, or do it with a shell script, as well. But, this is how I did it, and you're likely to run into all the same little quirks I ran into:
Interesting. I’ve been trying to use bwrap, slirp4netns, and mitmproxy to create a simple Python script to get save shell for development. But it’s a huge time sink (and I might resort to podman)
It blocks network access by default, mounts only what you specify, and you can add a customization layer. This is all done in the container itself (srt for network blocking). It doesn't implement a central point for secret sharing, MCP exposure, etc. So it might not have enough features for some but it works well for my needs.
I just found through this thread yoloai which has an apple container backend, so its quite similar using that. My main issue would be that network access is allowed by default. https://github.com/kstenerud/yoloai
Several other projects listed here use libkrun which is an alternate implementation that works with Mac's HVF. smolvm, microsandbox, podman (with likrun backend), gondolin.
To everyone sharing their favorite container-based sandboxing solution: Docker Sandbox does not use containers for isolation. It spawns the workload in a libkrun-based micro VM, which has vastly different security properties.
I got excited for this not because this didn't exist before, but because Docker putting their weight on this would imply a broader adoption and better integration in the industry. I am sad that they are asking for a login here though, which doesn't make any sense to me.
Not a substantive comment on content but hopefully constructive feedback on presentation:
Holy moly, on mobile I was trying to read the example console screenshots/snippets and then it would just unexpectedly change. Took me a little while to figure out it’s some kind of carousel for the examples, and not more screenshots/snippets loading and pushing down content (or me going crazy). Please don’t do this on mobile sites, just let me scroll through the examples!
Nono has been my daily driver since the start of the year. It's not a perfect sandbox -- that's for sure. For example, the default network rules let you escape via a global TMUX server. But it is extremely practical. It gives me enough guarantees to feel confident about running in YOLO mode. So far nothing has gone awry.
As you as your Go build fails because you haven't put the local cache dir in the "allowed directories", you'll understand how painful this is, as well as most tools based on bubblewrap/sandbox-exec. There is a difference between a clean environment with standard setup vs a layer on top of everyone's existing tools/setup, especially in a enterprise environment.
(I'm sure you can spend time to come up with a proper bubblewrap configuration that allows go build to succeed, but it's probably not worth the effort.)
Eg, if used with Colima in macOS, it means I can run a devcontainer in an isolated VM and Nono inside the devcontainer can restrict a lot what can and cannot be done.
You get credentials proxying and network outbound limits.
I just made my own devcontainer that I copy on any project and load whatever harness I want in that repo. Harnesss' config and auth are simply mounted from the host, so no setup required at all.
I don't get how it works, or works well -- presumably those domains are behind CDNs, and IP addresses are unpredictable. So the firewall allows traffic to specific IPs that are resolved at the time the script is run, but not after that? What if the same domain is resolved again without going through the cache, and it becomes a different IP?
And even if that works, this is a very short list. As soon as you reach for Go, Rust tooling etc nothing works. So you need to manually maintain this list which is nothing but painful trial and error.
I'm currently facing this issue. I've resorted to implementing my own execution environment albeit limited.
It goes like this:
- bash script parser + interpreter (with hooks for things like file open, execute etc.,)
- wasm executor for execution.
- wasm implementations of common tools like coreutils, grep, sed etc., from the uutils project.
- wasm implementation of python by a VMware backed project.
- entirely virtualized filesystem using Go's io/fs.FS. (tmp dirs can be implemented using any backend)
Works like a charm for the limited usecase I have. There are definitely some drawbacks with threading and especially with preopens in wasm. But a cheap sandbox for simple file explorations and minimal computations.
Operating systems ought to be providing us the utilities we need to safely sandbox processes (agent or otherwise), but they appear to not be interested in the job
I have a solution based on Nix that can be used to generate reproducible container images: https://github.com/nothingnesses/agent-images . It lets you customise which agents, harnesses, or any other packages you want included in the VM and it uses `agent-box` for sandboxing.
It's a batteries included alternative to firecracker with a couple of new ideas tossed into the mix i.e. portable like a container (bake into a single file and rehydrate the vm anywhere), dynamic resource allocation, etc.
Does this support Linux yet? When I previously looked it did not (the reason being that they were already using VMs on Windows/macOS but not on Linux). Every time I see an announcement I think "great, they must've added Linux now then", but the linked pages always have Windows + macOS instructions but not Linux.
All the open GH issues about supporting Linux that I subscribed to have gone unresponded to.
OpenShell looks like a good alternative, but it still has "Do not use in production" plastered all over the website, which doesn't fill me with confidence yet
I wrote a tool to use `bubblewrap` to containerize any agent (at least all the agents I've used a couple of times), and bind mount the system stuff read-only, so the agent has your "usual" environment, but they can only see the project. Their history persists (either through a bind mount or a "shadow" copy of the history that only the wrapped agent sees), the agent can still create and manage containers of its own using podman's rootless mode, etc. It's nearly instant to start because it's just a namespace (plus a few copied files for the container support and session history); no container needs to be built/fetched/updated/whatever. bubblewrap is extremely well-tested as it is used by flatpak and several other large projects, so I trust it quite a bit (more than I trust Docker).
bubblewrap may work well for you and your specific workflows/projects but not in an enterprise setting where everyone already has a different setup on the host and needs something different inside the container. It's impossible to deploy a solution like that with bubblewrap -- configuration itself is going to be a nightmare. Which is why Docker Sandbox is aimed at teams/enterprises.
I also hit the same issue recently. No Linux and no Windows on arm. AI sandboxing has a lot of options but none feel complete just yet. It's hard to commit to something, especially if reviewing tools to aide in company policies.
Regardless, I'm hoping something that isn't behind a login screen is going to win out.
Anyone serious about security will want to bring their own sandbox anyway, not trust these, often proprietary, agents. I've never run an agent outside a sandbox. My first bubblewrap script for `claude` is now over a year old. The tools are available and if you learn to use them you can run any program in a sandbox.
But, in any case, why put in effort doing something people don't expect or ask for? We can assume everyone running agents is either a) using their own sandbox, or b) doesn't care. I think we can guess which category most people fall into. You could maybe argue about responsibility, but I don't think you can argue about "serious engineering".
> Anyone serious about security will want to bring their own sandbox anyway
Exactly. It's not as though it's difficult. It never occurred to me to not do this from day one, and it astonishes me that anyone runs this stuff bare metal. Since then, I've brought several other people on board, and that's all they've ever seen: I don't think they'd know how to run outside a sandbox, and that's just fine.
I was all in on sandboxes and safehouse for my agents but the moment I got into iOS development it felt like my hand was forced to just run Claude / codex / pi directly on my machine because nothing else could do the dev loop.
It’s been a painful reality for me, I’m going against core pieces of how I feel I should be interacting with agent harnesses and yet, I need to get the work done so
Hey, I work at Docker and my team works on mcp integration with sbx. A solution I've been trying is this:
1) Enable the xcode mcp server: https://developer.apple.com/documentation/xcode/giving-exter...
2) Add the xcode mcp server to sbx: `sbx mcp add xcode --command xcrun --args mcpbridge`
3) When you create the sandbox, use `--static-mcp xcode`. For example: `sbx create --static-mcp xcode claude .`
Make sure you have at least v0.38.0 of sbx. This makes a bridge from inside the sandbox to the xcode tools on your host, so be aware that it can run whatever tools you give it on the host. But the agent itself is still sandboxed.
I’m sure they fixed this, but since everyone runs docker containers as root… is every file this thing writes going to be root owned? Does it have root access to any resource to give it visibility to?
Models start going to extreme, damaging lengths to achieve ambiguous prompts[0]. Having good sandboxes is now a must IMO.
But sbx is a bit annoying to use with OpenCode for instance (which has zero sandboxing by default, unlike codex CLI or Claude Code). You cannot easily change ~/.config/opencode/opencode.jsonc AFAIK.
We run OpenCode currently, but need to improve the setup for users, so would like understand more of the issues you have sandboxing it, if you can share
Can someone more versed in Docker explain to me how this is different than building my own docker container from a Dockerfile for using Pi agent harness? That's what I do currently. I use Docker Desktop in windows as the backend for that.
Docker containers use Linux kernel features to create an isolated environment, running on the same machine as docker is.
This creates a virtual machine, with its own kernel, and runs the container in there.
This gives stronger isolation and security guarantees.
I have the same question as GP. Your answer helps a little but not really. I might be naive, but I was under the impression that malicious code escaping a docker image and running amok on my host system was not something I should be too worried about. Especially if I run docker in rootless mode. Is that wrong?
For clarity I’m actually using podman, not Docker.
Oh no, you should definitely be worried about that. Podman might make it harder to escalate to host root, or manipulate other containers, but it is still vulnerable.
Now I'm curious to know how hardened the Docket Sandbox orchestration interface is. I guess we can assume they have run Mythos against it for a few weeks maybe? It's unclear.
Unsettling. I mean, is there any reasonable way to develop software in 2026? I've already sworn off ever installing npm directly on my host. Containerizing everything is laborious enough, but running a separate VM for everything?
You just need to work out the threat model for what you're working on. For trusted containerized workloads, where the attack surface is minimal, just containerization is fine. However, agents can do just about anything on your computer if you allow it and people aren't really shying away from `--dangerously-skip-permissions`, so better hardening (VMs, microVMs) is desirable.
Unless we're talking 0-day/CVE, running an unprivileged container is as trustable as a VM. The only difference is how strictly you want to hold the memory/CPU bar. Infact on linux, containers are more lightweight than VMs.
LLMs are great at finding 0-day, and people are rubbish at updating their containers and hosts to patch b-day.
Containers have access to the kernel ABI, and as shown in the latest kernel exploits, all the memory handling surface that exposes. The virtualisation interface, offering fewer services, is significantly harder.
Containers are obviously lighter than VMs, both to start and to schedule, but firecracker is pretty fast. gVisor pays overhead per syscall vs at startup.
Got it. 0 days are possible so throwaway containerization. You should blog about it, will help millions of developers and companies. Heck, even consult with the hyperscalers - they will be riddled with their workloads.
I know some people want to run their agents when their computer is off, but I imagine a solution like this will be much more common than paying for a remote sandbox (i.e on fly.io or exe.dev), especially because it'll be free.
Though, they need to remove the login requirement.
1. You can run sandboxes locally
2. You can control them securely over the internet, for when you're on the go
3. You can migrate them to cloud VMs if you want
If I can toot my own horn, I'm trying to build that :)
Still early and the local sandboxes are experimental right now
Before you use no sandbox at all use this or one the many similar projects but it's alway worth remembering that Docker is not a security boundary. It never has been meant to be and never will
become one.
cgroups are a mechanism designed for hierarchical organization and resource distribution. Against a malicious and capable actor, and that is how we have to treat AI agents, cgroups will not withstand.
Also, the kernal is an interface too big for what an AI agent needs and is therefore offering a gigantic attack surface completely unnecessarily.
Wouldn't it need a super critical exploit, I mean zero-day vulnerability, to escape from that kind of sandbox ? And if you think further, then isn't that risk also applicable to pretty much any kind of sandboxing ?
Container escapes are more common than you think. Common enough for AWS not to rely on containers for their serverless functions, common enough for Google to say: "Untrusted code shouldn't rely on the container security boundary [..]" [1]
The same is not applicable for any kind of sandboxing for two reasons:
1. The boundary is in the kernal’s own code, enforced by the thing you are trying to be protected from. -> Use a VM
2. The kernal is a gigantic attack surface -> Use gVisor
I tried Docker Sandboxes but last time I checked you could not configure custom volume mounts, making more complex setups impossible. For work I need two directories for context for the agent to have access to…
When starting a sandbox, you can specify the mountpoints you want. It just defaults to the current directory. You can also specify some of those mounts as read-only as well.
Of course, but that was not part of my workflow and I found it quite strange that this was simply not possible especially when docker-compose can easily do this
The sandboxing problem is perhaps the greatest justification for doing agent integration via existing human interfaces rather than low level shell access. Granting access to shell is a super obvious path (it's easy) so I can understand us wanting to fight for it. But we should consider the other paths as well before we make our final stand.
Automating browsers with LLM agents properly requires a lot more work than Process.Start into powershell, but the advantages can be immense once you have achieved integration this way. Incrementally maintaining this integration is generally easy because human users cannot tolerate rapid changes either.
It's a hell of a lot easier to convince management to adopt a robot that looks and acts like a human employee than one that looks like a combine harvester. The combine is far more efficient, but it is also totally indiscriminate. Nothing constrains its appetite except for the invisible fence imposed by GPS. The amount of infrastructure required to keep farm equipment from running astray is incredible. In the context of agriculture, the added complexity is definitely worth it. We don't want to have to recreate the same thing with our technology if it can be avoided. Sandboxes and security isolation boundaries are not things to aspire to. These are costs to be paid for admission to something more valuable.
I tried it and it worked great at first but I had multiple issues with it, the disk space usage was growing significantly, I need to login multiple times for each sandbox, it's closed source and not possible to customize to my need.
One other thing, I want to be able to handle multiple repos in the same sandbox and have a standard workflow around worktrees (one worktree per repo, all the worktree mounted in the VM).
I do not see any value proposition in this - if I need a sandbox, I make one with Dockerfile, Bubblewrap or virtualization. What I am missing? An enforced required login is a net negative value - it means rug pulls in the future.
I did try to use those for some stuff:
* Login requirement is something else
* It's closed source last I checked
* Pretty slow/unstable
There are many better namespace/container based options, VMs may be moderately more secure but when you more or less trust your agent and code you can do with lesser containment. And with the recent CVEs in kvm honestly there isn't a huge deal of difference vs namespaces.
(I'm building https://xbin.dev/ for some time now for managing my personal code/apps, a project which started specifically after Docker Sandboxes broke on me some time ago)
So Docker is finally adding native support for a microVM backend? I wonder how it well it will compare to using the Kata Containers 4 runtime with the KVM/QEMU backend.
This is nowhere near "exactly this". Docker Sandboxes uses micro VMs, you just use regular containers which have completely different security properties.
I'm confused:
1. If I run this on Mac, then inside the sandbox / microVM, am I still running MacOS or some Linux distribution?
2. If the only thing that's mounted from the host is the $PWD, how does it guarantee that it has all the system libraries that I have installed on my host system? e.g. my `/opt/homebrew` libraries or `sudo apt install libfoo-dev` headers
This is what I currently do, but my software uses docker and docker mounts act as a bypass for the file system restrictions, plus docker processes started outside the sandbox allow network proxy escape.
Currently, I don't allow the agent access to docker, start docker myself, and then do short-lived sandbox-free sessions when the agent needs to do things that interact directly with docker; but that's annoying.
So essentially you can get latest Pi/Node pulling from that image:
`sbx run -t ghcr.io/shaftoe/sbx-template-pi:latest shell`
Like others here I'm also saddened by the login requirement but at the moment this is the best UX I could find for running sandboxed agents, the "kit/mixin" concepts are neat and I make use of them too: https://github.com/shaftoe/sbx-template-pi#stacking-the-extr...
I use it (sbx), but I don't 100% trust that it actually works, and I would prefer something open source where the limits of the sandboxing could be tested and explored.
Maybe we should just ssh into separate development machines to ensure real and verifiable sandboxing? (as was totally standard before Docker became a thing)
Like many people, I suspect, I used Claude to write my own agent sandbox that suits my needs very well. Investing my time in a propietary product has become a hard sell.
I did the same thing. It was my first "vibecoded" project. I've been using it every day and it's great. I'm writing a custom Rust network stack for it right now. Gonna replace the current nftables firewall with it.
As for Docker Sandboxes, I'll just ask Sol literally right now to see what it does better than my virtdev, and then I'll improve virtdev instead of using Docker.
> Were you following any patterns/standards/advice on what you needed to protect against?
Just the general knowledge that sharing a kernel with untrusted software is too dangerous, that hardware virtualization is an infinitely smaller attack surface and that the entire industry will be in deep shit if people or AI breaks hypervisors.
Initial threat model was supply chain attacks but eventually grew to include AI harnesses as well. Not very worried about them hacking me, more about accident prevention.
So that means each VM must be running a completely independent kernel that's fully isolated from the host's file system. They must also have fail closed network filtering built in.
In summary, it's a QEMU VM orchestrator with a base OS image and project specific delta images. VM lifecycle is managed by systemd. System level isolation is already pretty good and it already solves the "AI wiped out my $HOME" problem. I'm currently working on a custom network stack to replace the nftables based firewall.
You want to prevent the agent/others from reaching your home directory and other things. As long as you don't mount/sync directories/files from/to the container, so no mounting like "-v $(pwd):/app", but instead copy in, then when done, copy out.
And of course, instead of doing the "copy in > copy out" process manually, get your local agent to write a bash script that does that for you, given what directory you're in, and you're basically G2G.
Sure, but all projects are version controlled? You only mount the project dir so you can only loose your current changes - which is the same if you copy...
The one thing I wonder about is how you enforce the usage of Docker Sandboxes vs running the agent on the host directly, apart from scanning machines for binaries
Doesn't everyone do this now? It's hardly a new idea. Yet every time someone proposes the idea, people fawn over it and proclaim it the best thing ever.
Yes, you can inject tokens via a proxy. What else is new?
Who is doing it as first class feature with at least adequate UX?
I have skimmed alternatives offered in comments to this post (vibepod-cli, code-on-incus, opencode-docker, sandboxy, smolvm, amazing-sandbox) and none of them seem to do credentials injection at the proxy level.
I wish they solved the issue happening for years on MacOS where Docker keeps up eating all available free space and ends up requiring restart of the whole machine, instead of Gordon and other useless shit.
I used this for a while then decided to build my own suites that pack individual harness and respective host state (config, plugins, skills, etc.) into an image. Works better and much flexible in my opinion.
Do the agents come preinstalled in the images? Or do they somehow use whatever I’ve installed locally? The former makes sense to me but then I’m wondering whether the sandbox images stay up to date with new releases of each image.
Ah let's see, do they still want you to LOGIN, in order to use a local dev tool? Yes, yes they do. No thanks Docker. You can keep your buzzword reasoning as to why this is needed.
I don't want this that bad. I want the agent to have open access to my system because it actually does important administrative things for me. It is THAT convenient and powerful.
Here's what I want: REALTIME OBSERVABILITY/POWERPOINT.
I don't want to just see what command it ran. I need graphics... what part of the file system it is touching, what network entities it is contacting. If it's running SQL I want the parsed query handed to me in a syntax highlighted and well formatted interface. Imagine that star trek computer presenting automated infographics while someone is doing a presentation, you know what I'm talking about? It's like a automated powerpoint as the agent does it's thing.
I need to understand my agent and what it typically does so I can dangerously wield it. I treat the agent like a gun in a live shooting scenario. That's how I want to use the LLM.
Sandboxes have their purpose. Just like how shooting ranges have their purposes. But I need to fire my gun in the real world and real world is a warzone.
TO me, that's the important distinction: sandboxing limits what the agent can do but it doesn't necessarily enforce that the agent must run inside the sandbox. You need a separate control layer to enforce that boundary.
You design the sandbox so the agent starts in that layer. The next thing you can do is to limit the network access, this is what I'm working on right now.
Not sure what you mean by "just". Containerisation is generally understood to mean something like what Docker does, which includes sandboxing but a whole lot more on top, like image management etc. Bubblewrap is just sandboxing without the rest of containerisation.
One of the demos I run is how easy it is to circumvent the harness limits. For example, I can configure a harness not to access file `secrets.txt`. But, then I can immediately have it create a Python file that can read any file and have it read `secrets.txt`.
At the end of the day, "please" isn't security. You want to know that the agent can only do and access the things it should access.
Just a small meta note: most of the comments in this thread appear to be posting their own codebase (typically AI-generated) that accomplishes the same goal. It's interesting that this problem is simultaneously in high demand and yet considered trivial enough to vibe code per-user solutions to it.
Customize the exact environment of your container from the ground up (harnesses, tools, base image, packages, mounts, etc) and enter with a single command.
One correction: this isn't containers. Each session is a microVM with its own kernel on the platform's native hypervisor: Hypervisor.framework, WHP, KVM. We wrote a new VMM (not Firecracker) to make it more effective across platforms.
Explained a bit more here about the architecture and why those choices were made: https://www.docker.com/blog/why-microvms-the-architecture-be...
I run it with superset and then each git worktree is mounted in a sandbox that is configured for each repo i work in.
Closest open source I have seen is https://earendil-works.github.io/gondolin but the DX is not as polished. https://exe.dev/ would be perfect but it does not come with outbound firewall.
Does anyone have a better alternative?
https://github.com/kstenerud/yoloai
Outbound firewall is `--network-isolated`: egress is denied except the agent's own API endpoints plus domains you allow, enforced sandbox-side (working on host-side enforcement now). `--network-none` if you want nothing.
Credential brokering works the way you describe (currently Claude-only, I'll add more as time allows). The API key stays on the host, a local proxy injects it into the outbound request, and the sandbox never holds anything worth stealing. Other agents' credentials currently arrive as read-only file mounts instead (weaker, and something I'll fix soon). Generalising the injector is the obvious next thing.
One difference from your setup: yoloAI copies your worktree instead of mounting it. The agent works on the copy, you `yoloai diff`, and `yoloai apply` replays the commits into your real repo. That's deliberate. Docker's own security docs talk about the dangers of bombs being left behind in a live-mounted dir (git hooks, package.json scripts, Makefiles, IDE task config), which diff/apply avoids.
Isolation is per-sandbox rather than fixed: runc, gVisor, or Kata VMs (QEMU or Firecracker) on Linux; Seatbelt or full macOS VMs via Tart on a Mac.
Cloudflare/artifact-fs does lazy shallow git clones with a FUSE filesystem. https://github.com/cloudflare/artifact-fs
Would that be faster?
Re: sandboxing methods like Clawk, Amla sandbox, bwrap, agentvm, ARM64 MTE with wasmtime-mte: https://news.ycombinator.com/item?id=48893850
A few months ago now I started adding seccomp sandboxing to jinja2rs and then liboverlayfs support to ansiblers (which are early Rust ports).
Haven't finished that, but I started working on a VM format that stores signed machine state into an OCI container repository, using the hypervisor migration support of KVM/QEMU.
Though this is not safe yet if ever, VM migrations are probably another way to sandbox and deploy en masse.
It also provides a nice TUI for network policy management and prebuilt sandboxes which have claude code, codex, etc.
I'm using it as my main driver since months.
Support for running agent harnesses in unprivileged podman containers is on my feature list. :-)
Only real alternative is to use microvms. My goto solution for this are apple/containers.
For the vast majority of developers, containers are enough, which is why they are ubiquitous while vms are less common. Ofc that ubiquity has led to lazy configuration, which is how the jailbreaking can occur. Knowing what you are doing with containers is a requirement to use containers as an AI sandbox.
Why microVMs? I never ever run a container, AI harness or other, in something else than a full on VM. I could use a microVM but in any case I really don't see why I'd run a container on one of my bare metal OS: the place of a container is inside a VM (or microVM).
Especially for AI harnesses where the threat of an escape is very real: the more defense in depth, the better.
And If I can use rootless Podman instead of "rootfull" Docker, the better. Most of my containers are Podman btw.
> My goto solution for this are apple/containers.
To each his own: my goto solution is an actual server on my LAN with shitload of cores and memory and plenty of scripts to provision VMs etc.
I really don't understand why people are YOLO'ing containers on their bare metal OS.
It has network filtering + placeholders for secrets.
OSS, no logins needed
(Not affiliated with them, just tried it out last week.)
Documention at this moment in an early stage.
Overall, its a great project but for me was simpler just use Virtual Machine Manager (libvirt GUI).
I wish all luck to the maintainers, but probably DX-wise I will prefer to have more granular or predictable controls (eg micro cloud from Canonical).
https://github.com/NVIDIA/openshell
We run cloud sandboxes, and have some experimental local sandbox support that is fully OSS.
Main thing for amika.dev is you can control the sandboxes and agents interchangeabley by SSH, web, or API, and can expose the services the agent is working on over signed URLs
Entire sandbox config is a TOML file
We're going to improve the local OSS sandbox mode and add better network controls over the next couple weeks.
Ultimately, what we're building kind of like if Tailscale and Firecracker had a baby, with a messaging protocol for remote controlling any sandboxed agent
It's free to try out. Still a lot to build, so we really appreciate any and all feedback about what we should focus on
I have a dedicated container for that. It can run its own Docker daemon and other system services if needed.
Apart from the Claude login token, it has no SSH keys or other credentials. I push everything I need to it from the local machine. And I pull the Claude generated outputs from it.
Of course, this kind of setup requires a stack which can run or at least be tested without any credentials.
What I don‘t have compared to sbx is an outbound firewall, but my VM does not have any personal/interesting data, only a vanilla Fedora installation and the project dir with open source code, so I do not care much about exfiltration.
https://GitHub.com/jgbrwn/vibebin
Has egress and ingress filtering, egress can be bound to host/internet/subnet or even better to internal apps (which are each separate netns) meaning you can do your own firewall/vpn/whatever per sandbox. Plus you control what other components in the sandbox env the app can communicate with.
Really not built for day-to-day dev work though, more like automating your company/life / getting rid of SaaS (e.g. for technical Founders / Sales etc, not exactly useful for dev work)
Internally uses a single VM + Incus containers, supports docker/Kubernetes in each sandbox, has integrated worktree management.
thanks
So I wrote a wrapper around Gondolin which allows me to do that and a few other things: https://github.com/codethief/tuor
(Warning: Still very much experimental / underdocumented.)
The definition for your cloud sandboxes is just a TOML config in your repo
We also have an API and CLI to let users message the agent from outside or across sandboxes
We're still building a lot, so if you have any time to try it out (amika.dev) and give feedback, that is worth gold to us!
It seems with gondoling i need to explain the agent to run commands in the sandbox, but then where does the agent run itself?
[0]: https://earendil-works.github.io/gondolin/workloads/
Their example implementation with pi uses a pi extension so that pi runs on the host but the read/write/bash/etc tools run in the guest. Doesn’t have to be that way though.
And that's just an easy case - really if it has any network access at all it can come up with a clever way to route a request through the network such that the key comes back somewhere in the request. If you scan for it inbound too, the machine can obfuscate it.
Our agents are trained to be so intensely helpful and they have such intricate knowledge of how things work that they will do some incredibly clever tricks to do what you ask them to do.
edit: You may want to look into tokenizing proxies as the general application of this concept.
There's also an ability to create kits where you can setup credential injection into other services as well.
https://github.com/pjlsergeant/byre -- slightly different security model, but lazer-focused on developer experience; my daily driver and I love it not just because I wrote it. The TUI is great for configuring and setting up instant boxes just how you want
https://pleasedonotescape.com/ -- a list of every other agent jail I could find, filterable by open-source and whatever else you want
Basically, it's a plain QEMU/KVM VM on a stock Debian cloud image: device model stripped down to a virtio disk, a virtio NIC and a serial console, nested virt off, no passwordless sudo in the guest. It also ships a containment check that scans outward from inside the guest, so the network boundary is something you can verify.
Wrapping the whole environment rather than a single agent session puts supply chain attacks inside the boundary too. A poisoned npm or PyPI package, or a compromised editor extension, lands in the VM instead of on the host. That was the original reason I set this up; agents just made it more urgent.
There's no per-domain egress allowlist; the policy is "internet yes, private addresses no". Secret injection isn't built in either, though Infisical's agent-vault on the host as an egress proxy covers that part.
Wrote the whole setup up here, in case it's useful:
https://karamatli.com/posts/network-isolated-kvm-sandbox-ai-...
> So rather than pick one, this post advocates layering both, in the spirit of defense in depth: a sandbox VM wraps your containers along with the whole toolchain, and that sandbox reaches the internet but has no route to anything private.
Yup it's the only proper way.
And that is true not just for AI harnesses/agents (that shall try to escape), but also for stuff like Plex/Jellyfin/Immich/private pastebin etc.
If you care about security, there really simply is zero reason to run containers on the bare metal.
I find solutions like this to be a like trying to patch a leaking boat on a lake with duct tape. It will help, but it's not a proper solution.
Also, often the tasks you want the AI to perform are in the outside world. Like "connect to my servers, and figure out X and Y".
The proper way is permission isolation. I run a small k8 cluster in the homelab and I have 3 types of pod/agent combinations for my AI agents. Read only, one that can change my gitops but it needs to create PRs that admin approves, and admin.
Likewise with code. I have a forgejo git instance where agents have ability to create feature branches and so on, but merging is gated.
Those things require "GH enterprise features".
In fact more and more things we do at home will require "enterprise features". Why? Because a person with AI is basically a small team, but some of team members behave like Chimps on crack... So security must be top notch.
https://www.mrafayaleem.com/blog/sandboxing-claude-cli-with-...
Both went through stuff the sandbox was set up to allow.
What's a "microVM" and what's the security model here compared to using real virtual machines with actual constraints on breakouts?
Is it marketing fluff?
Incus/LXD has had VM's for a long time now.
Firecracker is designed to start a VM in under 125ms and 5MB. Netbsd advertises that you can direct-boot a MICROVM kernel configuration in under 10ms.
https://en.wikipedia.org/wiki/Memory_ballooning
There is a need for a Docker desktop like GUI for this market.
The world didn't fit around me, so I made "me" fit around "the world". I bet those stubborn dinosaurs can learn a new trick or two also, if management lets them do it during working ours...
Also, WSL (Windows Subsystem for Linux) has been baked into Windows for a long time and makes it very easy to play with Linux, as does using the Hyper-V VM system. Any developer unfamiliar with Linux because they use Windows, has little excuse.
Modern .NET did not went cross platform by accident, and Java development has always been "develop on Windows deploy on UNIX", in corporations where Mac tends to have little presence.
I use something in between - a single Ubuntu VM, into which I spawn multiple Incus LXC containers for the agents. The containers only use 50 MB or so per instance (separate systemd, ...). This way I pay the VM RAM tax only once, and the agents are still contained inside the VM if they manage to escape the LXC containers.
This way I don't pay tens of VMs "tax" but basically only three (plus one or two VMs I use for testing enhancements to my VMs provisioning / optimization / securing setup, when I work on that).
I don't use LXC (I could) but regular containers, inside VMs.
I'll look into the stripped down "micro" VMs but then I don't spend my days launching VMs/shutting them down so it's not a big deal.
The VM sits on its own libvirt network in a dedicated firewall zone, and specific directories are shared via filesystem passthrough. To keep the agent from accessing anything related to Git, the actual gitdir is stored on a separate path outside the mount point.
I review the git diff manually and commit it from the host.
Then again, I guess GUI is a pain in Docker. I tend to operate through Zed and an ACP harness though, so my GUIs are sort of “inside the container” anyways.
Would suggest Hashicorp Packer or cloud-init for deterministic images, not hard to setup or use. LLM's have little problem with them either I find.
If you need quicker environment rebuilds consider using something smaller like alpine as the base, though once you setup a golden image even heavy things like debian are fine.
(it's in the doc, but kinda strange to not see some instructions on the main page, probably distro related)
https://github.com/swelljoe/flar
I’ve been using this pretty extensively for a few months on Mac and Linux and have been super happy with it.
> That runs the codex OCI in a qemu microvm.
AFAIU it's actually the other way around: krun spawns a libkrun-based (not QEMU-based) VM inside a crun container. Source: https://github.com/libkrun/libkrun/discussions/538#discussio...
So with your solution you get the additional security benefit of containerizing the hypervisor on the host.
https://docs.docker.com/ai/sandboxes/security/credentials/
It blocks network access by default, mounts only what you specify, and you can add a customization layer. This is all done in the container itself (srt for network blocking). It doesn't implement a central point for secret sharing, MCP exposure, etc. So it might not have enough features for some but it works well for my needs.
I just found through this thread yoloai which has an apple container backend, so its quite similar using that. My main issue would be that network access is allowed by default. https://github.com/kstenerud/yoloai
Several other projects listed here use libkrun which is an alternate implementation that works with Mac's HVF. smolvm, microsandbox, podman (with likrun backend), gondolin.
eg: https://josecastillolema.github.io/podman-wasm-libkrun/#libk...
Holy moly, on mobile I was trying to read the example console screenshots/snippets and then it would just unexpectedly change. Took me a little while to figure out it’s some kind of carousel for the examples, and not more screenshots/snippets loading and pushing down content (or me going crazy). Please don’t do this on mobile sites, just let me scroll through the examples!
Haven't tested it yet, but it seems to address the same issue as Docker Sandboxes, but in a different way.
But it’s not as well sandboxed for sure.
(I'm sure you can spend time to come up with a proper bubblewrap configuration that allows go build to succeed, but it's probably not worth the effort.)
Eg, if used with Colima in macOS, it means I can run a devcontainer in an isolated VM and Nono inside the devcontainer can restrict a lot what can and cannot be done.
You get credentials proxying and network outbound limits.
How is Docker Sandbox better sandboxed?
https://github.com/iodize6399/ai-devcontainer/tree/main/.dev...
I quite like the 'features' layer system, adding extra tools to container in a declarative plugin-like way
Being able to 'safely' run with skip permissions has been a gamechanger
I especially like the firewall it has.
And even if that works, this is a very short list. As soon as you reach for Go, Rust tooling etc nothing works. So you need to manually maintain this list which is nothing but painful trial and error.
It goes like this: - bash script parser + interpreter (with hooks for things like file open, execute etc.,) - wasm executor for execution. - wasm implementations of common tools like coreutils, grep, sed etc., from the uutils project. - wasm implementation of python by a VMware backed project. - entirely virtualized filesystem using Go's io/fs.FS. (tmp dirs can be implemented using any backend)
Works like a charm for the limited usecase I have. There are definitely some drawbacks with threading and especially with preopens in wasm. But a cheap sandbox for simple file explorations and minimal computations.
(I am testing this now as a backend for my pet project which currently supports firecracker and gvisor. No network.)
All the open GH issues about supporting Linux that I subscribed to have gone unresponded to.
OpenShell looks like a good alternative, but it still has "Do not use in production" plastered all over the website, which doesn't fill me with confidence yet
I wrote a tool to use `bubblewrap` to containerize any agent (at least all the agents I've used a couple of times), and bind mount the system stuff read-only, so the agent has your "usual" environment, but they can only see the project. Their history persists (either through a bind mount or a "shadow" copy of the history that only the wrapped agent sees), the agent can still create and manage containers of its own using podman's rootless mode, etc. It's nearly instant to start because it's just a namespace (plus a few copied files for the container support and session history); no container needs to be built/fetched/updated/whatever. bubblewrap is extremely well-tested as it is used by flatpak and several other large projects, so I trust it quite a bit (more than I trust Docker).
https://github.com/swelljoe/flar
Regardless, I'm hoping something that isn't behind a login screen is going to win out.
https://github.com/pkhamre/opencode-docker
But, in any case, why put in effort doing something people don't expect or ask for? We can assume everyone running agents is either a) using their own sandbox, or b) doesn't care. I think we can guess which category most people fall into. You could maybe argue about responsibility, but I don't think you can argue about "serious engineering".
Exactly. It's not as though it's difficult. It never occurred to me to not do this from day one, and it astonishes me that anyone runs this stuff bare metal. Since then, I've brought several other people on board, and that's all they've ever seen: I don't think they'd know how to run outside a sandbox, and that's just fine.
I was all in on sandboxes and safehouse for my agents but the moment I got into iOS development it felt like my hand was forced to just run Claude / codex / pi directly on my machine because nothing else could do the dev loop.
It’s been a painful reality for me, I’m going against core pieces of how I feel I should be interacting with agent harnesses and yet, I need to get the work done so
1) Enable the xcode mcp server: https://developer.apple.com/documentation/xcode/giving-exter... 2) Add the xcode mcp server to sbx: `sbx mcp add xcode --command xcrun --args mcpbridge` 3) When you create the sandbox, use `--static-mcp xcode`. For example: `sbx create --static-mcp xcode claude .`
Make sure you have at least v0.38.0 of sbx. This makes a bridge from inside the sandbox to the xcode tools on your host, so be aware that it can run whatever tools you give it on the host. But the agent itself is still sandboxed.
Also if your thing doesn't work with `pi` out of the box, then low effort
But sbx is a bit annoying to use with OpenCode for instance (which has zero sandboxing by default, unlike codex CLI or Claude Code). You cannot easily change ~/.config/opencode/opencode.jsonc AFAIK.
[0]: Black Hat OpenAI-Hugging Face incident: https://www.youtube.com/watch?v=87DyyMV0kCY&t=1021s
My startup (https://github.com/gofixpoint/amika) copies agent configs into local or cloud sandboxes
We run OpenCode currently, but need to improve the setup for users, so would like understand more of the issues you have sandboxing it, if you can share
Still obviously you should run all untrusted code in a sandbox, but extreme actions like that would be very unusual with the model that shipped.
For clarity I’m actually using podman, not Docker.
Now I'm curious to know how hardened the Docket Sandbox orchestration interface is. I guess we can assume they have run Mythos against it for a few weeks maybe? It's unclear.
Unless we're talking 0-day/CVE, running an unprivileged container is as trustable as a VM. The only difference is how strictly you want to hold the memory/CPU bar. Infact on linux, containers are more lightweight than VMs.
So yeah, not "vulnerable".
Containers have access to the kernel ABI, and as shown in the latest kernel exploits, all the memory handling surface that exposes. The virtualisation interface, offering fewer services, is significantly harder.
Containers are obviously lighter than VMs, both to start and to schedule, but firecracker is pretty fast. gVisor pays overhead per syscall vs at startup.
So yeah, more vulnerable.
On Linux, you can run docker directly on the host, but you can also very easily setup a vm with incus and run docker from there.
Though, they need to remove the login requirement.
Still early and the local sandboxes are experimental right now
https://github.com/gofixpoint/amika
cgroups are a mechanism designed for hierarchical organization and resource distribution. Against a malicious and capable actor, and that is how we have to treat AI agents, cgroups will not withstand.
Also, the kernal is an interface too big for what an AI agent needs and is therefore offering a gigantic attack surface completely unnecessarily.
This is not my main point though. Both are based on cgroups and cgroups are the wrong tool for the job.
The same is not applicable for any kind of sandboxing for two reasons:
1. The boundary is in the kernal’s own code, enforced by the thing you are trying to be protected from. -> Use a VM
2. The kernal is a gigantic attack surface -> Use gVisor
[1] https://docs.cloud.google.com/kubernetes-engine/docs/resourc...
Say I want to use pi inside a container. If I wrap pi within a bubblewrap or within nono, how is that less secure than using a vm?
Also, I think most people run containers inside VMs anyway and not directly on their hosts (on Mac and windows you have to use a vm anyway).
It doesn’t change the fact a malicious process is still attacking the same kernel , but it can reduce what it can do to that vm.
Example: sbx run claude ./ ../another-project:ro
Automating browsers with LLM agents properly requires a lot more work than Process.Start into powershell, but the advantages can be immense once you have achieved integration this way. Incrementally maintaining this integration is generally easy because human users cannot tolerate rapid changes either.
It's a hell of a lot easier to convince management to adopt a robot that looks and acts like a human employee than one that looks like a combine harvester. The combine is far more efficient, but it is also totally indiscriminate. Nothing constrains its appetite except for the invisible fence imposed by GPS. The amount of infrastructure required to keep farm equipment from running astray is incredible. In the context of agriculture, the added complexity is definitely worth it. We don't want to have to recreate the same thing with our technology if it can be avoided. Sandboxes and security isolation boundaries are not things to aspire to. These are costs to be paid for admission to something more valuable.
One other thing, I want to be able to handle multiple repos in the same sandbox and have a standard workflow around worktrees (one worktree per repo, all the worktree mounted in the VM).
These were some of the reasons that led me to build: Clawk - https://github.com/clawkwork/clawk
https://apps.apple.com/app/aifcc-ai-first-computer/id6782364...
als has lots of agents + and typical dev packages (node tooling, python tooling, ....) preinstalled
'su agent'
'curl domain/install.sh | sh'
'runagent'
There is no reason to require a login for creating local mini sandboxes.
If you’re on Apple, native solutions like “container-machine init” come built in and are pretty good, if you’ll only be on Apple hardware.
There are many better namespace/container based options, VMs may be moderately more secure but when you more or less trust your agent and code you can do with lesser containment. And with the recent CVEs in kvm honestly there isn't a huge deal of difference vs namespaces.
(I'm building https://xbin.dev/ for some time now for managing my personal code/apps, a project which started specifically after Docker Sandboxes broke on me some time ago)
Like:
$ podman run -it --rm -v .:/workspace local-dev-ia /usr/bin/oc
Configured with a .env file. Hope to do it hopefully before the end of the week.
This is nowhere near "exactly this". Docker Sandboxes uses micro VMs, you just use regular containers which have completely different security properties.
Docker is always a pain to use and this way I don't have to re-install everything a billion times for every different project.
Currently, I don't allow the agent access to docker, start docker myself, and then do short-lived sandbox-free sessions when the agent needs to do things that interact directly with docker; but that's annoying.
Ideally, they should run in _different_ sandboxes.
The environment might corrode the harness (e.g. rogue npm/pip packet would manipulate agent harness config).
On another topic, can't help but notice that "leading coding agents" somehow does not include Pi.
To work around that limitation I came up with this https://github.com/shaftoe/sbx-template-pi
So essentially you can get latest Pi/Node pulling from that image:
`sbx run -t ghcr.io/shaftoe/sbx-template-pi:latest shell`
Like others here I'm also saddened by the login requirement but at the moment this is the best UX I could find for running sandboxed agents, the "kit/mixin" concepts are neat and I make use of them too: https://github.com/shaftoe/sbx-template-pi#stacking-the-extr...
Maybe we should just ssh into separate development machines to ensure real and verifiable sandboxing? (as was totally standard before Docker became a thing)
The option left is to use SSH to sign commits which is a no-go for a different reason.
I use it regularly to run Claude/Codex with permission checks disabled.
https://github.com/cvhariharan/mvm
As for Docker Sandboxes, I'll just ask Sol literally right now to see what it does better than my virtdev, and then I'll improve virtdev instead of using Docker.
Just the general knowledge that sharing a kernel with untrusted software is too dangerous, that hardware virtualization is an infinitely smaller attack surface and that the entire industry will be in deep shit if people or AI breaks hypervisors.
Initial threat model was supply chain attacks but eventually grew to include AI harnesses as well. Not very worried about them hacking me, more about accident prevention.
So that means each VM must be running a completely independent kernel that's fully isolated from the host's file system. They must also have fail closed network filtering built in.
> Anything you can point the rest of us to?
I have published my virtdev's design document.
https://github.com/matheusmoreira/virtdev/blob/master/DESIGN...
Yes, it is AI generated.
In summary, it's a QEMU VM orchestrator with a base OS image and project specific delta images. VM lifecycle is managed by systemd. System level isolation is already pretty good and it already solves the "AI wiped out my $HOME" problem. I'm currently working on a custom network stack to replace the nftables based firewall.
And of course, instead of doing the "copy in > copy out" process manually, get your local agent to write a bash script that does that for you, given what directory you're in, and you're basically G2G.
Start by mounting just your repo and passing in the keys for the agent. Take it from there, it's like software engineering, you iterate.
When you run into issues you expand the tools in the container available to it.
Better sandboxing for AI agents is exactly the main reason for containers improvements on macOS and Windows, with a few talks at WWDC, and BUILD.
Not sure how much they would get from Linux users then.
Yes, you can inject tokens via a proxy. What else is new?
I have skimmed alternatives offered in comments to this post (vibepod-cli, code-on-incus, opencode-docker, sandboxy, smolvm, amazing-sandbox) and none of them seem to do credentials injection at the proxy level.
Also fnox now does credentials proxying.
For those who do not trust
AND do not want to use some other, free VM for some reason?container run --rm -it -v "$(pwd)":/work -w /work myaiimage /bin/bash
Other than the login problem, it’s a decent option.
https://engine.build/lab/agent-sandboxes
The open source section specifically.
Here's what I want: REALTIME OBSERVABILITY/POWERPOINT.
I don't want to just see what command it ran. I need graphics... what part of the file system it is touching, what network entities it is contacting. If it's running SQL I want the parsed query handed to me in a syntax highlighted and well formatted interface. Imagine that star trek computer presenting automated infographics while someone is doing a presentation, you know what I'm talking about? It's like a automated powerpoint as the agent does it's thing.
I need to understand my agent and what it typically does so I can dangerously wield it. I treat the agent like a gun in a live shooting scenario. That's how I want to use the LLM.
Sandboxes have their purpose. Just like how shooting ranges have their purposes. But I need to fire my gun in the real world and real world is a warzone.
Or do you mean something else?
Don't give it shell access, just predefined tools.
One of the demos I run is how easy it is to circumvent the harness limits. For example, I can configure a harness not to access file `secrets.txt`. But, then I can immediately have it create a Python file that can read any file and have it read `secrets.txt`.
At the end of the day, "please" isn't security. You want to know that the agent can only do and access the things it should access.
The other url is their marketing page.
Yes, Linux is supported.
[1] https://github.com/apple/container