Use case

Run untrusted code off your laptop.

Execute LLM-generated scripts on throwaway isolated computers with egress policies, then discard the machine when the job is done.

Why this is hard

Model output is code you did not write

It is usually fine. "Usually" is not a security posture when the thing running it has your AWS credentials in its environment.

Production is the worst place to run it

The environment that holds your database credentials, internal APIs, and customer data is not where you want model-generated code executing unchecked.

A throwaway run still needs real tools

Isolation does not mean a stripped-down container. The agent still needs your stack, your packages, and controlled egress to do useful work without touching what matters.

How a risky run is contained

  1. It gets a throwaway computer

    Isolated, disposable, and holding nothing that matters. Damage is bounded by what you deliberately put in it.

  2. Outbound traffic goes through the gateway

    You allow the specific hosts the job needs. Tokens are injected at egress, so the code never holds credentials and cannot reach what you did not put on the allowlist.

  3. Credentials stay outside the sandbox

    If the run needs an API, the gateway injects the token at egress. The code cannot read a secret it never receives.

  4. Every call is logged, then it is destroyed

    You keep the audit trail of what it tried to reach and throw the machine away.

What you get back

  • Run output
  • Denied-request log
  • Per-request audit trail
  • Destroyed environment

The blast radius

One disposable computer per run

No shared filesystem or network with production or other tenants.

Egress allowlisted at the gateway

You name the hosts the job needs. Tokens inject at egress so the code never holds credentials; everything else is denied and logged.

Full audit trail, then the machine is destroyed

Every outbound attempt is recorded before the environment is thrown away.

Who runs this

Each employee can be many agents, working in Salesforce, running web searches, and more. When an agent writes Python for analytics or a heavy calculation, we dispatch it to Islo and run the code in a sandbox. That is where our story began.
Barel Tayouri, CTO · Stealth AI startup

Where teams go next

Common questions

Why shouldn't I run LLM-generated code on my laptop?

Generated scripts may be wrong or malicious. Islo runs them on disposable computers with scoped egress instead of your laptop or production network.

Can I send Python from an LLM response straight to Islo?

Yes. Pipe generated code into an isolated computer through the CLI or SDK and collect stdout, stderr, and artifacts.

Is the computer deleted after the script finishes?

You can stop and remove the computer, or snapshot state if you need to inspect results later.

Can the generated script access the internet?

Only through the egress rules you configure. Gateway policies can allow required hosts and block everything else without placing credentials inside the computer.

Can I inspect stdout, errors, and generated files afterward?

Yes. Collect stdout, stderr, and artifacts through the CLI or SDK, and keep or snapshot the computer when you need a longer investigation.

Put this workflow on autopilot.