cd /news/ai-tools/sandboxing-with-minimal-effort · home topics ai-tools article
[ARTICLE · art-137878] src=yorickpeterse.com ↗ pub= topic=ai-tools verified=true sentiment=· neutral

Sandboxing with minimal effort

Yorick Peterse merged a new Inko feature that lets applications sandbox themselves with minimal effort, using Linux Landlock, macOS Seatbelt via sandbox_init, FreeBSD Capsicum, and OpenBSD pledge and unveil. The cross-platform API handles platform-specific details such as setting Landlock rules for the ELF program interpreter /lib64/ld-linux-x86-64.so.2, but on FreeBSD the sandbox is a no-op because Capsicum requires fundamentally restructuring a program. Peterse noted that memory safety alone is insufficient since code is still written by developers, and argued that using an LLM to write code makes things worse.

read4 min views1 publishedSep 23, 2026

The other day I merged a new feature for Inko that I think is quite interesting: the ability to sandbox an application with minimal effort.

While memory safety is a goal of Inko (ignoring the usual escape hatches such as the FFI), memory safety only gets you so far. Most notably, the code is still written by developers and developers are, by and large stupid, myself included. And no, using an LLM to do the writing instead doesn't improve things; if anything it makes it even worse given the average LLM has the intellect of a talking parrot with a bad drinking habit.

One approach popularized by Docker is to run the program in a container. Not just because it makes distribution easier, but also because additional restrictions may be applied to the container, such as limiting the files it has access to. For example, this website is served by shost, a static file server written in Inko. To run the server I use the following Podman quadlet:

[Container]
Image=ghcr.io/yorickpeterse/shost:main
Pull=missing
ContainerName=shost
Exec=shost --sites /var/lib/shost/sites --tls /var/lib/shost/tls --log json --ip ::
PodmanArgs=--memory 1024m
ReloadSignal=HUP
Network=host

ReadOnly=true
UserNS=auto:uidmapping=0:500,gidmapping=0:500,size=1024
DropCapability=all
AddCapability=CAP_NET_BIND_SERVICE

Volume=/var/lib/shost:/var/lib/shost:z,ro

[Service]
Restart=on-failure
RestartSec=60
TimeoutStopSec=15

[Install]
WantedBy=default.target

If you're not familiar with quadlets, they're essentially systemd unit files for running containers using Podman. It's a bit similar to Docker Compose, but a lot nicer to work with.

Either way, the point here is that in the above quadlet I apply some restrictions to the container: all capabilities except for the "bind" capability are dropped, and the files that need to be served are mounted into the container as a read-only volume. Oh and if you're wondering what that UserNS line is for, that's to work around this issue.

Now this is great and all, but it would be even better if the application itself included some mechanism to restrict its own capabilities, regardless of how it's run.

Fortunately, most mainstream operating systems offer some way for an application to sandbox itself. For example, on Linux one can use Landlock while on macOS one can use Seatbelt through sandbox_init. FreeBSD in turn has Capsicum, and OpenBSD has pledge and unveil.

The sandboxing API provided by Inko uses these primitives to provide a cross-platform way of sandboxing your application, and tries to handle platform specific behavior/differences as much as possible. For example, on macOS allowing a file to be executed is easy but when using Landlock you also have to set up the appropriate rules for the ELF program interpreter (/lib64/ld-linux-x86-64.so.2 in most cases). If shared libraries are in a non-standard location you also need to make sure those can be read.

Of course this new API is not without its trade-offs. Most notably, on FreeBSD the sandbox is a no-op. Not because I was too lazy to make use of Capsicum, but rather because Capsicum requires you to fundamentally change the structure of your program. On Linux and macOS you can apply sandbox restrictions without having to change your program, other than whatever lines of code are necessary to list the sandbox rules (i.e. the above enable_sandbox method). Capsicum on the other hand works a little differently: once you call cap_enter you can no longer open resources using the usual system calls such as open. Instead, Capsicum requires that you either open all the appropriate resources before calling cap_enter, or that you open directories ahead of time and then use openat to open resources relative to that directory. In some cases you may also have to use libcasper. Of course for simple programs this may not be much of an issue, but for larger programs it may require you to extensively change how they are written. openat itself also has its issues.

That's not to say you can't make Capsicum work or that it's somehow "bad", rather it means you (unfortunately) can't use Capsicum in many instances unless you're willing to adjust your program to specifically cater towards FreeBSD and Capsicum.

So how difficult is it to sandbox an Inko application using this new API? Well, here's all that was necessary to sandbox shost:

import std.sandbox (Sandbox, bind, read)


fn enable_sandbox(config: ref Config) {
  let s = Sandbox.new

  match config.tls {
    case Some(v) -> s.path(v, read)
    case _ -> {}
  }

  s.path(config.sites.path, read)
  s.tcp(config.port, bind)
  s.enable
}

That is: we allow access to the directory containing TLS certificates (if TLS is enabled), we allow access to the directory containing the files to serve, and we allow binding to the TCP port the server listens on (e.g. 443 when TLS is enabled). Everything else is denied.

While one could consider applying a sandbox to something like shost redundant, given it already runs in a restricted container, it's also so easy to use this new API there's no reason not to use it.

And that brings me to the following that's worth repeating before we wrap things up for the day: the value of a security feature lies not in what it can do, but rather in it's ease of use. I think the API provided by Inko does a pretty good job at achieving just that, though I may be biased on account of, well, having written it.

── more in #ai-tools 4 stories · sorted by recency
── more on @inko 3 stories trending now
sponsored brought to you by zahid.host 4,200+ EU-deployed projects
reading about agents? ship yours in a single git push.

Run your AI side-project on zahid.host

EU-based hosting, git-push deploys, automatic HTTPS, no cold starts. Free tier with a custom domain — perfect for shipping the agent you just read about.

$git push zahid main
Live at https://your-agent.zahid.host
Get free account → Pricing
from €0/mo · no card required
LIVE [news/sandboxing-with-mini…] indexed:0 read:4min 2026-09-23 ·