In a post dated 19 August 2026, Simon Willison described tasking Claude Fable 5, running in Claude Code for web, with evaluating smolmachines.com as a secure sandbox for running untrusted Python and JavaScript code. The stated goal was to limit RAM and CPU usage (protecting against infinite loops such as 'while true'), block network access, and restrict filesystem access to designated files only — with a view to executing user-provided tasks such as data transformations. According to Willison, the agent 'quickly ran into a problem': the Claude Code for web environment could not run smol machines. He wrote that the agent then switched to 'Plan B', installing smolvm and running the tests directly in a GitHub Actions runner against that branch. Willison characterized this as 'a creative solution to the environmental limits posed by Claude Code for web' and 'another example of Fable being relentlessly proactive.' The post does not report detailed results of whether smolvm fully met the security and resource-limiting requirements. This account comes from a single source, Willison's own blog, and reflects one exploratory experiment rather than a formal review.
- Willison used Claude Fable 5 in Claude Code for web to evaluate smolmachines/smolvm as a sandbox, posted 19 August 2026.
- The goal was running untrusted Python/JavaScript with RAM/CPU limits, no network access, and restricted filesystem access.
- The agent could not run smolmachines in Claude Code for web and instead installed smolvm in a GitHub Actions runner.
What it means for you
If you ever plan to let users or an AI run code you didn't write — say, to transform uploaded data — you need a 'sandbox' (an isolated box that limits how much memory and time the code can use and blocks it from touching the network or your files). This post is one developer's early exploration of a tool called smolvm for that job. It's a signpost, not a verdict — it doesn't tell you whether the tool actually works well.
Try this
If you're building something that runs untrusted code, read Willison's post and note smolvm as a candidate to test yourself — but run your own trials before relying on it.
Who should care
Developers and technical founders building apps that execute code supplied by users or AI agents.
Skip this if
You don't write software or run code from untrusted sources — this has no bearing on using ChatGPT or off-the-shelf AI tools.
Sources: Simon Willison — read the original