AI News
All stories
Official

Prime Radiant releases three external sandbox servers for coding agent code

Prime Radiant has released three open sandbox servers that run agent-generated code outside the agent's own environment, with a shared localhost API.

Safety & Security · · 3 min read

Prime Radiant has released three open sandbox servers that run code written by coding agents outside the agent's environment. The agent stays on the host, and only the commands it asks to run go into a separate Linux guest. The company calls this the external sandbox pattern. Simon Willison's post on the Prime Radiant blog says the approach is meant to make it safer to try local models, where a mistake or a prompt injection attack could cause damaging actions on a machine.

The three servers differ mainly in how they isolate the guest. external-sandbox-v86 is a Node.js server that runs 32-bit Alpine Linux inside the v86 x86 emulator, compiled to WebAssembly. Its README says the npm package bundles the guest image, about 36 MB including the Linux image. external-sandbox-container is a Python server that runs Alpine in an Apple Containers instance, with two CPUs, 512 MiB of memory and --network none. external-sandbox-quicksand is also a Python server, built on Microsoft's Quicksand library, which ships QEMU and the Alpine image as Python wheels. All three share one localhost protocol, so the same agent client can talk to any of them, on ports 8686, 8687 and 8688 respectively so they can run side by side.

The post argues that the alternative, where the agent runs inside the sandbox, lets the agent modify its own environment. It also says that keeping the boundary outside the agent makes it easier to audit. The pattern is contrasted with the model-based safety checks in Claude Code's auto mode and Codex's auto-review. The authors say those checks appear to work well but still depend on a model judging each action correctly. A sandbox, they write, is a protection that does not depend on any model's judgment.

The integration differs by agent. Claude Code's Bash tool can be routed through a shell prefix, but its native Read and Edit tools still run on the host. The post notes that the mount rules limit only what Bash commands can touch. Codex connects through its exec-server protocol, which the post says is experimental. The authors tested against Codex CLI 0.164.0 and warn that other versions may need checking. Pi works through an extension that replaces its bash, read, write, edit, grep, find and ls tools. The post says that running Pi with --no-builtin-tools means nothing runs on the user's computer if the extension fails to load.

Network access is off by default. The proxy is an allow-list of exact origins, with optional methods and paths, and it can substitute a placeholder header with a real credential held by the host server. The post's example uses GITHUB_TOKEN, and the guest only ever sees the placeholder. The post warns that an allowed host could still echo a credential back in its response, so credential rules should be used only with trusted services and kept narrow. The post also says the sandboxes are experimental prototypes that have not been audited. A read-write mount still lets the agent damage anything in that folder, so the folder should be under version control.

Some details limit how far the design goes. The v86 README says the API is meant for trusted local processes and has no token authentication, so any other local process can call it. The container README says Apple Container is the only backend it implements, and that its isolation differs from the v86 prototype. It does not claim that every operation an agent performs is contained.

What changed

Prime Radiant has released three open sandbox servers that run coding agent code outside the agent's own environment, using a shared localhost protocol on ports 8686, 8687 and 8688.

Type: New capability · Applies to: external-sandbox-v86, external-sandbox-container, external-sandbox-quicksand

Why it matters

Teams that want to test local models or run agent-written code can isolate execution from the host machine, reducing the damage a mistake or prompt injection could cause. The network allow-list and credential substitution address two common risks in agent workflows.

What you can do

Engineering leads evaluating local models for coding agents could trial one of the sandbox servers in a non-production environment and check its proxy and credential handling against their own policies.

Original source

Further reporting

Where we saw it

1 source

Simon Willison

MicrosoftPrime Radiant

AI wrote this summary from the sources listed, and nobody reviewed it individually before publication. Check the original source before you rely on it. How it works