Prime Radiant and Simon Willison released external sandboxes that run Claude Code, Codex and Pi commands inside an isolated Linux VM
A sandbox limits what an agent's mistakes can reach without relying on any model's judgment, and a new open project makes that easy to try.
What happened
Prime Radiant, working with Simon Willison, released three prototype sandbox servers for coding agents. Each runs Alpine Linux in isolation: one in v86, an x86 emulator compiled to WebAssembly and run from Node.js, one in Apple Containers, and one in a QEMU virtual machine through Microsoft's Quicksand library. All three expose the same localhost API, and the team published a specification and a conformance test tool so others can build compatible servers.
They call it the external sandbox pattern. The agent runs on your machine as usual, and the commands it executes go to the sandbox. The sandbox sees only the folders you mount, read-only or read-write. It has no network access by default, and an optional proxy allows only the hosts you list. The proxy also does credential replacement, adding API keys to requests for specific hosts so the secret never enters the sandbox or reaches the agent. A web UI lets you change the rules and review a log of network requests.
Coverage differs by agent. Pi sends all of its file and shell tools to the sandbox, and Codex sends its process and filesystem operations. Claude Code sends only Bash commands, so its native Read and Edit tools still act on the host. The authors argue that approving every action by hand leads to decision fatigue, and that cheaper or local models make more mistakes and fall for prompt injection more easily.
My take
The credential replacement proxy is the piece I'd look at first. Many agent setups hand the agent an environment full of API keys, and any prompt injection that reaches a shell can read them. Keeping the key in a proxy the agent can't inspect, and limiting outbound hosts, removes the two things an injected instruction needs most.
The Claude Code caveat matters. If your threat model includes the agent reading files it shouldn't, Bash isolation alone won't cover it, so keep secrets out of the project folder as well. These are also prototypes, and Codex support relies on an experimental protocol, so I'd use them for evaluating cheaper or local models on client code before putting them in front of anything in production.
More posts
- Open OCR models now read 1,000 pages for under $0.10, and better serving saved more than switching modelsNov 20, 2026
- Gemini 4 Neon costs $3 in and $15 out per million tokens and tops Zapier's AutomationBench at 56.3%, though it uses more output tokens per taskNov 20, 2026
- Gemini API's Tool Retrieval loads tool schemas only when the model asks for them, so agents with hundreds of tools send fewer input tokensNov 20, 2026
