curiosity

Whitepaper · Change control

An AI admin you can approve

Sudo, the admin assistant in Curiosity Studio, changes your configuration in a sandbox, apart from your live deployment. Nothing built there runs until a person approves the diff. Written for the people who decide.

For CIOs and security leads Chapters 6, and a glossary Pages 21 Date October 2026

What it covers

Sudo does the work. A person keeps the say.

An AI assistant with admin rights is a risk when it holds the host’s authority, acts unreviewed or sees past its user. Sudo holds none of that. Every change waits as a diff until a person approves it, on the record.

  1. Guard 01

    Sandbox

    Nothing dangerous inside

  2. Guard 02

    Approval

    Nothing runs before

  3. Guard 03

    Permissions

    Sees what you see

  4. Guard 04

    Record

    One commit per approval

Chapters

Six chapters, from the fear to the questions

01 · Page 03All four

The risk is the authority, not the model

Three fears, and what each one is really about.

02 · Page 06Sandbox

Dangerous things are absent, not blocked

A sandbox with nothing in it to misuse.

03 · Page 09Approval

Nothing runs until a person approves

The loop, and an approval bound to one diff.

04 · Page 12Permissions

Sudo sees what you see, and no more

Your permissions, never elevated ones.

05 · Page 14Record

Every change is on the record

One tracked commit per approval, by name.

06 · Page 16All four

Questions to ask any AI admin

Seven questions, and three steps to try it.

The full paper

Ask Sudo. Approve what gets built.

Download the PDF

Your data. Your infrastructure.

Connected knowledge for AI systems