Security
How Saphan Studio keeps AI coding agents inside the boundaries you set, what it can prove afterwards, and what it does not cover.
Saphan Studio dispatches AI coding agents to machines you own. This section is the security model: who and what is trusted, how each run is confined, what is recorded, and — stated as plainly as the rest — what this release does not protect against.
Start here
| you are | read |
|---|---|
| a buyer or a manager | In one paragraph, What the risk actually is, then Documented limits |
| a security officer or an auditor | Design principles, the controls below, The record and Documented limits |
| an administrator | Administrator's runbook and Worked examples |
The controls
| control | what it does |
|---|---|
| Identity and PKI | who may sign what, and how a signature is checked |
| Machines | how a machine joins the fleet and proves it is the machine it claims to be |
| Capability | what each machine is trusted to be able to do, by measurement |
| Seats | separating agents that share a machine |
| Runners | the boundary each run gets on the machine |
| Sandboxing | which agent is confined where, and how that was established |
| Egress control | what a run may connect to |
| Guardrails | a checkpoint on the conversation between an agent and its model |
| Session recording | what is recorded of every run |
| Secrets | how credentials are referenced, never stored |
| The record | what you can prove after the fact |
Agent security: the whole set puts every control between an agent session and your machine on one page.
Agent security: the whole set
Every control that stands between an agent session and your machine — confinement, tool policy, egress, credentials, and the human gate — and where each one is documented.
Who this document is for
Two readers: an evaluator deciding whether to let agents run on their machines, and the administrator who operates the controls.