Simon Willison ran a test to see if the smolmachines framework could safely execute untrusted Python and JavaScript code within a web environment. He assigned the task to Claude Fable 5 running in Claude Code for web, asking the model to verify limits on CPU time and memory usage while blocking network and filesystem access. The AI immediately identified that the current container lacked the required virtualisation hardware flags, specifically noting the absence of /dev/kvm and CPU virtualisation extensions. This prevented the direct execution of smolvm machines inside the browser-based interface. To bypass this restriction, the model generated a GitHub Actions workflow using Ubuntu runners that do expose the necessary kernel interfaces. The system then ran the full test battery on that branch, collected the logs, and prepared to remove the temporary workflow before the final commit.
This approach demonstrates a practical method for overcoming environmental constraints when testing AI agents that require specific hardware capabilities. It shows how an AI can autonomously design infrastructure changes to validate its own operational limits without human intervention. The solution relies on standard open source tools rather than proprietary workarounds.
- Test environment was Linux 6.18.5-fc-v20 running as a Firecracker guest
- Initial failure was due to missing vmx and svm CPU flags
- Resolution involved a temporary GitHub Actions runner with /dev/kvm access
