---
title: 'In the News: November 17, 2026 (Morning)'
description: "Simon Willison and Prime Radiant release external sandboxes that run coding agents' commands in Linux, offline by default."
canonical_url: 'https://darkfactory.dev/news/2026-11-17-morning'
markdown_url: 'https://darkfactory.dev/news/2026-11-17-morning.md'
collection: news
date_published: '2026-11-17T07:05:00-05:00'
date_modified: '2026-11-17T07:05:00-05:00'
---

# In the News: November 17, 2026 (Morning)


Simon Willison has published a working pattern for containing coding agents without approving every command: the agent stays on your machine, and everything it executes goes to a Linux sandbox that sees only the folders and hosts you allow. It comes as three interchangeable sandboxes, one protocol, and setup for Claude Code, Codex and Pi. Claude Code users get the weakest version of the boundary, and the post says so plainly.

## 1. Prime Radiant ships "external sandboxes": the agent stays outside, its commands run inside

**[External sandboxes for coding agents](https://blog.primeradiant.com/external-sandboxes-for-coding-agents/)** · Simon Willison · Prime Radiant blog, November 16, 2026

The release is three prototype sandboxes that expose the same localhost API: one runs Alpine Linux in v86, an x86 emulator compiled to WebAssembly and hosted in Node.js; one wraps Apple Containers; one runs a QEMU virtual machine through Microsoft's Quicksand. A specification and a conformance testing tool for that API ship alongside them in the [external-sandbox repository](https://github.com/prime-radiant-inc/external-sandbox) (Apache-2.0). The sandbox mounts only the folders you name, read-only or read-write, and has no network access by default. An optional proxy allows HTTP and HTTPS to listed hosts and methods, and can perform credential replacement: code in the sandbox sends a placeholder token and the proxy swaps in the real key from the host's environment, so the agent never sees it. Willison rejects per-action approval as the alternative because it "quickly leads to decision fatigue where we instinctively approve everything." His case for the sandbox is one line: "A sandbox is a protection that doesn't depend on any model's judgment."

The three agents connect differently, and coverage differs with them. Pi routes bash, read, write, edit, grep, find and ls to the sandbox. Codex routes its process and filesystem operations, including `apply_patch`, through its experimental exec-server protocol; the team's automated tests pass against Codex CLI 0.164.0. Claude Code routes only its Bash tool, through `CLAUDE_CODE_SHELL_PREFIX`; its native Read and Edit tools still run on the host. The post lists its own limits: the sandboxes are experimental and "have not been audited" for sensitive workloads, a read-write mount still lets the agent damage that folder, and an allowed host could echo a credential back in a response.

**Why it matters:** The pattern moves the safety boundary out of the model and into a process outside the agent. The post argues this matters most for cheaper or local models, which it calls more prone to mistakes and easier victims of prompt injection. Coverage is uneven, though. Under Claude Code the sandbox limits what shell commands can touch, but the agent can still read and edit files on the host, so a team choosing this setup for containment gets a much wider boundary from Pi or Codex than from Claude Code.
