Run untrusted code like it's already compromised.
We're building isol8r so the risky parts of development, like installing packages, trying out a repo or letting an AI agent run commands, happen in a VM where nothing leaves unless you allow it, and you can see everything it tried to reach. All without slowing you down.
- allowedapi.anthropic.com:443in the analysis profile
- blockedpypi.org:443only allowed in the setup profile
- refusedDNS lookup for a7f3e9.collector.examplename not in the profile, so no answer
- blocked203.0.113.40:443address no allowed name resolved to
- blocked198.51.100.7:22not in the profile
- blockedregistry.npmjs.org:443the install task ended, so its access did too
The problem isn't new. It has got much worse.
Developers have always run code they didn't write, on laptops that hold SSH keys, cloud credentials, tokens and access to internal systems.
What's changed is the scale. Frontier AI models make attackers faster and more capable. At the same time, we've all started handing AI agents the keys to our machines: they install packages, run commands and read whatever is in front of them, and we give them more access every month.
Prompt injection makes it worse again. A single README, issue or web page with hidden instructions can turn your own agent against you.
Untrusted code is sometimes called toxic nuclear waste. That's the right attitude for nearly everything we touch during development:
- third-party libraries and their install scripts, even from vendors you trust
- scripts copied from the internet
- networks you don't control
- any code or content that could carry a prompt injection
Security scanners are reactive. They flag a malicious package hours or days after it's published, which is too late if you installed it in that window, and they can't recognise a zero-day at all.
So stop asking whether a piece of code is safe, and ask how much damage it can do when it isn't.
Where we're taking isol8r
Security that gets in the way gets switched off. isol8r is for engineers' day-to-day work, not production deployments, and the safe way of working has to be the easy one. The goal is to lower risk without costing you productivity, not to build a hardened production platform.
It feels like working on your own machine
Your editor, your tools and your agents work the way they always have. Setting up a safe workspace should take minutes, not a day of configuration.
Getting access is quick, not a ticket
When a task needs more, such as a registry to install dependencies, asking for it is one step, and you can see straight away what was granted and why something was blocked.
Untrusted work happens in a VM, not on your laptop
Assume whatever runs inside will be compromised, and keep your real credentials out of reach.
Nothing leaves unless you allow it
All network access is blocked by default. You allow only what each task needs, such as a package registry while installing dependencies or one API while an agent runs, and access goes back to nothing when the task ends.
DNS is a leak too
A lookup can carry data out by itself, so DNS gets the same default-deny rules as every other connection.
You can see everything
What the code tried to reach, what was allowed, what was blocked, and which process asked for more access and why.
A bad dependency should be a contained mess, not a leaked set of credentials.