Skip to content
Rykon AI Help

Safety and sandboxes

What a run can and cannot reach

Simple answers come straight from the control plane, which never runs code. Anything that runs code, a shell or files gets a fresh, locked-down sandbox whose network reaches one place, the relay. Your keys never enter it.

Updated 1 min read

A run answered directly works in the control plane, where no tool runs code. Anything that needs code, a shell, files or packages gets a new sandbox made for that task, kept only for the same task’s next sandbox step, and removed after it. The limits are in the system, not in the agent’s instructions, so nothing an agent reads or decides can lift them.

The sandbox hasThe sandbox does not have
A fresh filesystem with a read-only root and scratch spaceAnything from another task, or anything that survives the task
A non-root user, no capabilities, fixed memory, processor and process limitsA way to raise its own privileges
A deadline, after which it is stopped and removedA way to keep running past it
One credential, its run token, working only for this run, while it runs, from this sandboxYour account’s gateway key, your API keys, or anyone else’s
A network that reaches the relayThe internet, cloud metadata, our databases, or other runs

The relay

The relay is the only address a sandbox can reach. It checks the run token, forwards model calls to our AI gateway with your account’s sealed key (which it never hands to the sandbox), meters every call, takes the transcript and the result, and passes tool calls to the control plane to be decided.

Before a run starts work

Network rules are applied as a sandbox starts, which can take a moment. A run starts no work until it has checked that addresses it must not reach do not answer, and it refuses to start if they ever do.

What a run can and cannot reach · Help · Rykon AI