Simon Willison and Prime Radiant recently released three external sandbox implementations that move coding agents’ code execution off the host machine and into isolated environments. For enterprises, this is not an edge security feature; it is a threshold that determines whether agents can enter real business systems.

What This Is

External sandboxes—running untrusted code in a separate environment—use an “agent outside, sandbox inside” architecture. Developers can let Pi, Claude Code, or Codex execute code through a unified sandbox interface, while configuring network access allowlists and credential substitution. This reduces the risks posed by code errors and prompt injection attacks, which induce agents to execute dangerous instructions.

The project currently provides three implementations: external-sandbox-v86, based on WebAssembly; external-sandbox-container, based on Apple Containers; and external-sandbox-quicksand, which combines QEMU, Alpine Linux, and Microsoft’s Quicksand library. All isolate code execution while allowing the system to control the scope of external communications.

Industry View

We believe this shows that the coding-agent competition is shifting from “how much code can it write” to “can it execute code safely.” As agents gain access to terminals, file systems, and networks, enterprise buyers will care more about permission boundaries, audit logs, and credential isolation than generation quality and speed alone.

The risks are equally clear: a sandbox is not a universal guarantee. External services may still be misconfigured, while containers and virtualization layers may contain vulnerabilities. Overly broad network allowlists or abused replacement credentials can still cause damage. Adding sandboxes also increases deployment, monitoring, and permission-management costs. Small teams may not need a full solution, but once an agent touches production environments, security isolation should not be treated as optional.

Impact on Regular People

For enterprise IT: Enterprises can pilot coding agents with greater confidence while retaining control over network access and credentials. Before a pilot, however, they still need to specify which directories, interfaces, and data the agent may access.

For individual professionals: When developers use tools such as Claude Code and Codex, the scope of operations can be isolated, reducing the consequences of accidental file deletion or dangerous command execution. Permission requests and result reviews, however, remain essential.

For the consumer market: As tools make secure execution a configurable capability, agent features are more likely to appear in office software and developer services. Consumers will ultimately see less accidental interference and more stable automation, rather than the sandbox itself.