dataqbs

External sandboxes for coding agents

· Source: Simon Willison

Simon Willison explains how he and Prime Radiant have created a method called “external sandbox” to safely run code generated by AI agents. In this approach the agent stays outside the isolated environment, while any instruction that requires execution is sent to a dedicated sandbox server. The project offers three distinct implementations: a Node.js version that uses v86 to run Alpine Linux inside WebAssembly; a Python server that employs Apple Containers; and a third that combines Microsoft’s Quicksand library with QEMU and Alpine Linux, also accessible via Python. All share a local API that lets models such as Pi, Claude Code or Codex delegate code execution to the sandbox. Additionally, the environments provide a whitelist proxy for HTTP/HTTPS and a credential‑substitution mechanism, allowing the agent to make authenticated calls without exposing secrets. The interface lets users adjust the network policy while the sandbox is active.

This development is significant because it enables experimentation with local models without compromising the host system’s security, a crucial step toward broader use of generative agents in production and development settings.

Read the original article on Simon Willison

This summary is an informational synthesis produced by dataqbs.com. All rights to the original content belong to its author and the cited media outlet. We act solely as curators of technology news and claim no authorship.

Read this in Español · Deutsch · Français · Italiano