I like that your readme has a couple paragraphs comparing to popular tools in this space. Personally I feel it is a bit light on the security differentiators (if any). For example you are using the same fundamental primitives that are used in nsjail, runc, etc. Thus it seems you have recreated those libraries in some aspects, so would be curious to hear your rationale for approaching this way vs building explicitly on those primitives.
Anyway, my decision, for which I also evaluated the use of bubblewrap as a building block, was to err on the side of flexibility that calling fine-grained Linux APIs directly give. For a project like Drop, runc could be seen as very coarse-grained JSON-based API to Linux sandboxing calls (basically a single call: setup a sandbox, here is a json config that describes it), similarly bubblewrap is a coarse grained command-line API to Linux sandboxing calls. Reusing such tried and proved layers of course also has significant advantages, so as in case of many engineering decision, it wasn't super obvious which path is better.
Drop eventually integrated gVisor's runsc (as an option), which is also OCI-compatible Container Runtime, but this is to add a user-space kernel isolation layer.
At a glance, this seems to be providing most of what I was originally looking for when I ended up deciding I'd have to write it myself, but focusing specifically on Linux and providing a more full-fledged sandbox rather than only caring about a small set of permissions that I personally had a need for. Probably the biggest (and least hardened) feature that I spent time on in mine was trying to figure out how to allow arbitrary GUI apps so that I could run agents in it via Zed.
I'm definitely going to try this out and see how well it works for me. It's insane to me that this is something none of the big AI companies have bothered solving this yet other than via opaque rules built into their harnesses or absolutely awful manual rules that expect me to hard-code shapes of shell commands that I want to allow or not allow.
That's what Codex does out of the box, and it's not good against malware - i.e. a rogue npm packet (or even just codex after prompt injection) can read your ssh key and send it to the attacker.
As I said, opaque rules built into the harness rub me the the wrong way. They could change in an update without anything making it clear. Plus, I don't use Codex outside of work (I don't have any active paid subscriptions LLM offerings).
> it's not good against malware - i.e. a rogue npm packet (or even just codex after prompt injection) can read your ssh key and send it to the attacker
Ignoring the repeated references to software I don't personally use, I never said I was trying to hedge against malware. The use case for me is when I'm running agents directly based off of prompts that I give them and asking them to modify some files. If I wanted a solution for running code I didn't trust, I wouldn't rely on what I wrote, because that's not the intended use case at all.
I went with Docker because history has taught me that new isolation strategies _will_ have escape bugs at some point, and I’m distrustful that the LLM can’t find one if it wants to.
> Probably the biggest (and least hardened) feature that I spent time on in mine was trying to figure out how to allow arbitrary GUI apps so that I could run agents in it via Zed.
I actually have code for this if you want to use/fork/borrow it. TLDR, mine pretends to be an ACP agent so you run Zed on the host, but under the hood that binary is just creating a Docker container with your image and agent, copying files, etc, and then proxying ACP messages via web socket back and forth. Except for the built-in ACP read/write file and shell endpoints. Those get executed inside the container by the proxy by default, though there’s a config option to pass either or both through to the host.
Repo is https://github.com/SethCurry/abyss and the code you’d want would be in ‘internal/websockets/wsacp’.
It does wrap them in protobuf and there’s some router-like stuff so you can add your own non-ACP messages between the two ends of the proxy.
main also has experimental wasm plugin support in the proxies so you can block prompts/tool calls/whatever in an agent-independent way, or add RAG that works for every agent in the world, or whatever.
For me, it nails the convenience vs isolation aspect quite well, and I would like to get to a point where I can use it for all my development by default.
The main challenges that I still have, which none of the solutions I've found so far resolve, are:
- development of containerized applications, where e.g., docker/podman compose is used to bring up services
- development of GUI applications with hardware acceleration, like games (maybe security contexts for pipewire and wayland, like what's done in flatpak would be a good path?)
When I did my research in this space last time, https://litterbox.work/ was another option I've considered, and it is very nice too, but the friction is a bit higher than drop: it's slower to rebuild envs, lacks a "base" config, etc. Drop was much easier to adopt for me so far.
Based on past experience I'm guessing everybody is using Kubernetes, but what are the details? Stripped down VMs on Kubernetes? Hardened containers on Kubernetes? Can the rest of us spin up something comparable on K3s?
Anybody knows the answers and can share?
I am working on a project similar in spirit that uses microsandbox (libkrun) to run inside a tiny and fast VM. It includes other security properties that are needed for some workloads.
* network allow lists
* credential masking
* github allow list
There are going to be cases where you want to white list an org or a repo for read access that is not under your control and Github filtering will be a simple way to do that.
Github can be a source of hostile code, prompt injections, and exfiltration- you may want to lock down using repos that aren't yours.
The tool inherited this feature from the prior implementation and its something I am still exploring.
I use an alias to launch a Podman container with opencode in $PWD. It is fully ephemeral aside from the directories I map to it (usually a couple configs). XDG_HOME is local to that working dir. My only frustration is that my image is too minimal but that can be resolved.
Compared to your setup:
* /usr is from your host, so you don't need to maintain a separate image to have programs that you already have installed.
* username, hostname, your current directory and home dir paths are preserved in the sandbox (within a Podman container a home dir is /root)
* environment variables are easy to carry into the sandbox.
* environments are explicit (`drop ls` lists them) and can be removed with `drop rm`, so you don't need to track in which dirs you have started Podman if you want to cleanup XDG_HOME files.
It is likely that your Podman wrapper also solves some of these or they are non-issues for your. If your setup works well, I wouldn't switch to something different.
My own approach has the following features, hopefully that gets stolen into an actual softawre package at some point:
- Generate bwrap options (for use in bwrap or gVisor) rather than build the sandbox itself. I wont't trust a random sandbox.
- UDS based automatic HTTP proxy with popup or allowlist and reverse proxy to serve apps. (I won't trust any more net access).
- File based configuration (executable with shebang).
- Composable profiles and command line options for ad-hoc sandboxes.
- Manage system directories (use host's, use another distro's, use an OCI image, etc.) as well as home directory for the sandbox and how the "work" directory is shared.
Drop appears to do a few similar things, and it seems to fit well with how I use such systems. I would like to see more details about how the sandboxes are executed and about the security model.
As the container escapes with K8s shows, it is super tricky to get isolation right. E.g. what happens if a file you think is safe to write to is suddenly is replaced by one that isn’t.
The first Drop version (and the current default runtime) uses Linux namespaces alone to achieve this.
Then, gVisor was added as the second runtime, because it could be done in a way which is completely seamless from the user perspective, both runtimes produce identically looking sandboxes.
A lightweight VM could potentially be a third runtime, but I'm not yet sure it is possible to use a VM in such a way, that the sandbox is configured identically to the first two runtimes. Basically, quickly boot a kernel using the distro already in / and mount the same dirs Drop mounts with two other runtimes. An obvious problem I can already see is that /etc is not fully readable to the user running Drop, so using it to boot a VM will require working around lack of config file access.
Anyway, if VM support was possible, it would obviously have advantage of being fully compatible standard Linux kernel while providing very good isolation of the host kernel. gVisor does have some compatibility issue, as it is re-implementation of the kernel in Go.
Another advantage would be that with VM runtime it would be possible to start containers from Drop, which currently, due to issues related to nested namespaces, is not supported.
I agree. One thing I'm doing is to review past security problems in popular sandbox and container related projects and check if they apply to Drop. Drop is also rootless only (it won't even start as root), which helps to cut some classes of problems.
Can you explain to me what the benefit of using this is versus something like an LXC container? It is currently my preferred "temp dev env" that behaves like a full system. It can also be customized with cloud-init to load up my preferred settings.
It does not use process namespaces, and can run on Android (on Termux with proot-distro).
Drop is interesting because it is thinking about the approach differently - basically an extension of the user's existing userspace. This prevents files from being inadvertently deleted by the agent, but it doesn't prevent the agent from uploading files that it shouldn't be.
If you want to prevent a damage from a malicious dependency or a prompt injection, you need more robust protection (for example read-only everywhere exposes your ssh keys). You can build this on top of bwrap, but because it is a low level sandbox building block, you would likely end up creating some higher level abstraction on top of bwrap (for example srt and Flatpak are build on top of bwrap).
Drop is an attempt to create such a generic, high level sandboxing tool. I personally prefer to run agents already within a sandbox than to rely on a coding agent runtime to sandbox itself. Especially that by doing so, I can use the same sandboxing tool and config for installing other programs that need isolation, not just for running agents.
It isn't obvious to me what this does that we haven't been able to do for some time now.
It's isn't obvious to me what Docker does that we haven't been able to do for some time now.
Since it isn't obvious to your, Docker and friends also have an entire content management system (pulling, pushing, tagging images and layers) as well as a build system.
Can you make any similar guarantees about Drop?
BUT IT DOESNT WORK ON MAC OS :( bummer (it's Linux-specific)
Guess i'll test it on a VPS
I did a quick look-around last month and decided to start using bwrap. But manually configuring it on a per-project basis is irritating. So I rolled out something for my own use (+ a couple of friends) based on bwrap.
What I do:
- start with `bwrap --clearenv --unshare-all --die-with-parent --tmpfs / ...`
- every single file and folder and envar I need has to be mapped in. I have profiles in TOML, and `prepare/probe` commands to make this task simpler
- `--tmpfs /` means sandbox inits as `/home/user` on tmpfs unless you specify your own `home` and `user` keys.
- `--unshare-all` means there is no network inside the sandbox. So I used `socat` to run a HTTP/S proxy inside. Lets me control exactly which host+port combinations can be accessed. But this means nothing except HTTP(S) works. So no ICMP/UDP/TCP.
- profiles can be extended via extend syntax (otherwise you have to prepare/probe/manually specify everything per profile which is a nightmare). Which lets me do a base -> net -> coding chain.
All of the options I find seem to say something along the lines of, “We’re totally secure so long as you don’t Fizz the Bar or Twiddle the Quanzipulator”.
What is the most bullet proof, idiot proof option I should use?
"They worked in Symantec’s Threat Intelligence Lab in Culver City, the cyber equivalent of a biodefense lab, where researchers could unleash malevolent code on a “red” network—a sandboxed system air-gapped from Symantec’s business network—to observe its hostile behavior in a controlled environment. To reach the ground-floor lab, workers passed through several sets of security doors, each with progressively more restrictive rules. The final gateway kept all but a handful of workers out and physically isolated the red network from computers connected to the outside internet. Portable media were prohibited here—no DVDs, CD-ROMs, or USB flash drives were allowed —to prevent workers from mindlessly slipping one into an infested machine and inadvertently carrying it out of the lab with a malicious specimen stowed away on it."