Using smolvm as a sandbox for untrusted Python and JavaScript

An AI coding agent named Claude Fable was tasked with testing the smolmachines tool known as smolvm to see how well it works as a secure, fast sandbox for running untrusted Python and JavaScript code. The goal was to execute user-provided tasks, such as data transformations, while restricting RAM usage, CPU time, network access, and limiting filesystem access strictly to designated files.

During the initial phase, the agent encountered a roadblock because the Claude Code web environment lacked nested virtualization and access to a KVM device, causing the smolvm runtime to fail. To overcome this limitation, the agent devised a creative workaround by using GitHub Actions runners, which do expose the required hardware capabilities through dev kvm.

The agent successfully installed smolvm and ran a full battery of tests directly inside temporary GitHub Actions workflows on the target branch before collecting the logs and removing the workflow files. This experiment demonstrated how an AI agent can proactively navigate complex environmental constraints to evaluate specialized sandboxing infrastructure.