How the Django Co‑founder Used Claude to Reverse‑Engineer Its Own Sandbox
Simon Willison instructed Claude Cowork to explore its own application bundle, revealing that the AI runs inside a full Ubuntu 22.04 virtual machine on macOS via Apple’s Virtualization.framework, protected by a dual sandbox of VM isolation and Bubblewrap, with strict user and syscall restrictions.
Anthropic announced that Claude Cowork is now available to all Pro users. Shortly after, Django co‑founder and AI blogger Simon Willison conducted a "nest‑doll" experiment by directing Claude Code (the engine behind Cowork) to investigate its own sandbox.
"Nest‑doll" reverse: Claude, tell me who you are
Willison gave Claude the command:
"Dig around in the /Applications/Claude.app directory to figure out how the sandbox works."
He then prompted Claude to "Write a detailed report about the Linux container environment you are running in." Claude complied and produced a comprehensive technical report.
Revealed: Cowork runs inside an Ubuntu virtual machine
The report shows that Claude Cowork does not use a lightweight Docker container; it launches a full Ubuntu 22.04 LTS (Jammy Jellyfish) VM on macOS using Apple’s Virtualization.framework, which runs an ARM64 Linux system.
Kernel: Linux 6.8.0
CPU: 4 ARM64 cores
Memory: ~3.8 GiB
Disk: 10 GB root partition + 10 GB session partition
Dual sandbox protection
To prevent AI misbehavior, Anthropic applies two layers of isolation:
VM isolation: All operations occur inside the Ubuntu VM, separate from the host macOS.
Bubblewrap (bwrap) sandbox: Inside the VM, the Claude process is further confined by Bubblewrap, creating an independent network namespace, PID namespace, and a "die‑with‑parent" mechanism that destroys the container when the parent exits.
Strict permission controls
Running user: a non‑root account named brave-loving-maxwell.
Seccomp filtering: a tight syscall whitelist blocks dangerous operations.
Network proxy: all HTTP/SOCKS traffic is forced through a local proxy at localhost:3128, ensuring every network request is monitored.
This experiment gives a concrete view of the security model behind local AI assistants.
Security first, performance second
Anthropic chose a heavyweight VM approach to guarantee absolute safety: even if prompt‑injection or malicious code occurs, the damage is confined to the temporary VM, leaving the host macOS untouched.
Different sandbox strategies
Anthropic: Bubblewrap + full VM – most robust for a local client.
Vercel: Firecracker microVMs – lightweight, suited for high‑concurrency cloud workloads.
Cloudflare: Durable Objects / Isolates – V8‑based isolation, fastest startup but more limited.
Claude’s choice of a full Linux environment explains its ability to run Python, Node.js, and various CLI tools directly, effectively giving users a complete Linux computer.
Signed-in readers can open the original source through BestHub's protected redirect.
This article has been distilled and summarized from source material, then republished for learning and reference. If you believe it infringes your rights, please contactand we will review it promptly.
How this landed with the community
Was this worth your time?
0 Comments
Thoughtful readers leave field notes, pushback, and hard-won operational detail here.
