Simon Willison open-sources external sandboxes for coding agents
11/18/2026 — 11/18, 07:17·1 sources (all)·1 reports (all)
Key takeaways
- Simon Willison, working with research lab Prime Radiant, proposed an "external sandbox" pattern: code an agent wants to run executes in an i…
Background & analysis
On 17 November 2026, Simon Willison published a post titled "External sandboxes for coding agents" on Link Blog, describing exploratory work carried out in collaboration with Prime Radiant, the AI research lab behind projects including Superpowers. The post sets out what the authors call the "external sandbox" pattern: code that an agent needs to run executes inside a separate sandbox, while the agent itself stays outside of it. The stated motivation is to make it safer to experiment with local models, without the worry that a model mistake or a prompt injection attack could lead to destructive actions on the machine.
On 18 November 2026 at 06:52, Willison reported that the work had moved from a conceptual proposal to a set of released implementations. The team open sourced three sandbox service implementations that share a common localhost API. With this API, tools such as Pi, Claude Code or Codex can forward their code execution requests to the sandbox rather than running code in their own environment. The releases also include an optional HTTP/HTTPS whitelist proxy that can be enabled alongside the sandbox.
As of the available reporting, the three implementations have been published in open source form, with the localhost API and the optional proxy described as the current building blocks of the pattern. No further developments or follow-up announcements appear in the existing material, so the work is currently presented at the stage of open sourced reference implementations rather than a later milestone.
AI-generated from 1 reports · updated 25 minutes ago
- Heat index
- 98
- All sources
- 1
- All reports
- 1
- First seen
- 54 minutes ago
Reports on this story headlines open the original
Simon Willison and AI research lab Prime Radiant introduced the external sandbox pattern, in which code an agent executes runs inside a dedicated sandbox while the agent itself stays outside, reducing the risk that a local model's mistake or a prompt injection damages the machine. They also released three sandbox server implementations sharing one localhost API, configurable for Pi, Claude Code or Codex, with an optional HTTP/HTTPS allow-list proxy.
Simon WillisonAI score 63
Other stories people are talking about
- Heat index 1364NewGPT-6.2 Astra and Sol Now Available on ChatGPT, API and Perplexity15 sources
- Heat index 1268Nvidia board establishes national security and export control committee24 sources
- Heat index 1130SurgeNvidia chairman confirms no plan to bring Jensen Huang back as CEO13 sources
- Heat index 998California Sues OpenAI Over Agent Intrusion Into Hugging Face19 sources
- Heat index 952RisingAnthropic nears trading debut as Amazon and Google to hold nearly 30% combined stake12 sources
How is heat calculated?About the methodHide
Heat counts how many independent sources covered a story in the last 48 hours: one source counts once no matter how many posts it published, decaying with a 24-hour half-life. What ranks first is what many people are talking about.
This page aggregates public feeds. Headlines and summaries are machine-organized and remain the property of the original authors; verify important facts at the source. If you believe a headline or summary infringes your rights, email the contact address in the footer with the page URL and basis of your claim, and we will remove or replace it after review.
- Surge
- Discussion rising fast
- New
- First report within 6 hours
- Rising
- Still gathering discussion

