← All projects
Lock Down the Machine: Workstation Security for Claude Codemaster · ~190 min

The Boundary You Assumed You Had — Measure What Each Claude Code Sandbox Actually Contains

You have almost certainly told yourself a sentence like "it is fine, it runs in a sandbox." Which sandbox?

And what, exactly, does that sandbox stop? This project replaces that sentence with a measurement: a small, harmless instrument you build once, run inside every isolation boundary available on your own machine, and read as a single table. By the end you will know — not assume — which of your boundaries keeps a runaway command away from your credentials, which one only covers shell commands, which one your own configuration quietly widened, and which class of work has no business running on this machine at all.

🤝 Want this built for you? Start a request — our AI scopes it, estimates a price range, and matches you with a vetted AI Advisor.
Share this:
The Boundary You Assumed You Had — Measure What Each Claude Code Sandbox Actually Contains

🔒 Ambassador-exclusive build

Unlock this walkthrough

Two ways to open it

Just this build
$40 one-time

Unlocks this full walkthrough on your account — no subscription, and no account needed until you check out. Advisor-grade build.

Best value Rovvi AI Advisor membership
$69 / month

Every Ambassador-exclusive build, first access to new ones, and your spot on the advisor-only needs board. Or $300/year.

Become an Advisor to unlock →

What you'll do

A reusable, inert reachability probe plus a recorded blast-radius ledger that shows, boundary by boundary on your own machine, exactly which synthetic canary paths were readable and writable, which network destinations were reachable, and whether the SSH agent and host process table were visible — measured on the bare host, inside the built-in Bash sandbox, from a hook in that same session, through a whole-process sandbox runtime, inside a custom container in both its offline and mounted-and-networked profiles, and inside a shareable dev container built from Anthropic's published feature — together with a written routing policy that assigns each class of work to a boundary you have actually measured, one-word launchers that make the policy the easy path, a re-measurement check that exits non-zero when a boundary becomes weaker than the recorded ledger, and a completed drill in which you deliberately widened one boundary, watched the check catch it, and closed it.