$ cat wiki/entities/prime-radiant.md
Prime Radiant
Latest
- 2026-11-16
the article presents the three prototypes and integrations with Claude Code, Codex and Pi. It explicitly says the prototypes have not been audited su…
Overview
Prime Radiant publishes experimental tools for running coding-agent commands inside external Linux sandboxes while the agent remains outside. Its 2026-11-16 article describes the design and its limitations; this is maintainer evidence, not an independent security audit. (source)
Key People
Simon Willison authors the reviewed article. It does not establish a leadership roster or his employment status. (source)
Models & Products
- external-sandbox-v86: Node.js and WebAssembly run Alpine Linux through v86. (source)
- external-sandbox-container: a Python server wraps Apple Containers. (source)
- external-sandbox-quicksand: a Python server runs QEMU and Alpine Linux through Microsoft's Quicksand. (source)
- The prototypes share a localhost API, with a published protocol and conformance tool. Agent integrations route different sets of operations; see Agent Runtime Containment for the boundary details. (source)
Recent Activity
- 2026-11-16: the article presents the three prototypes and integrations with Claude Code, Codex and Pi. It explicitly says the prototypes have not been audited sufficiently for sensitive workloads. (source)
Strategic Position
Interpretation: separating the agent from command execution makes the execution environment replaceable across agents and sandbox implementations. Compatibility of the interface is distinct from assurance about every permitted action; writable mounted data remains exposed to modification. (source)
Related
- Agent Runtime Containment — tool routing, network controls and residual risks. (source)
- Agents (LLM Agents) — the broader execution workflow. (source)