Defender Kit Medium For AI Eng, AppSec, CISO AI-assisted
External sandbox pattern isolates agent-executed code from the host machine
Simon Willison and Prime Radiant released three sandbox server implementations for running code executed by coding agents. The agent runs outside the sandbox while the executed code runs inside a dedicated sandbox. The sandboxes include an optional HTTP/HTTPS allow-list proxy and a credential replacement mechanism.
Why it matters
Teams running local models or coding agents face risk from mistakes or prompt injection causing damaging actions on a host machine. Isolating agent-executed code limits what an attacker or faulty output can reach.
Do this
- Evaluate the external sandbox pattern for coding agents that execute code on developer machines.
- Restrict agent-executed code to an allow-list of hosts using an egress proxy.
- Keep real credentials out of the agent process and use a credential replacement mechanism where available.
- Review the sandbox server configuration before connecting Claude Code, Codex, or Pi to it.
Signals
- NoCISA KEV
- 21/100Priority
- 60%Confidence
Evidence
“external-sandbox-v86 runs a Node.js server which then uses v86 to run Alpine Linux inside WebAssembly.”
“The sandboxes also include an optional HTTP/HTTPS allow-list proxy, so you can control exactly which hosts your agent's code is allowed to communicate with”
Sources
Affected / related
external-sandbox-v86, external-sandbox-container, external-sandbox-quicksand, Claude Code, Codex, Prime Radiant, prompt injection
Published 2026-11-17 22:52 UTC · updated 2026-11-17 23:05 UTC · AI-assisted summary, checked against quoted sources · How we score