Trust and privacy
How Growth OX checks its work
The layers that sit between an AI model and your outcome: validators, release gates, approvals and undo, release signing and the data-flow audit. Each layer names the file it rests on.
The layers#
- Validators and never-allowed rules check an action before it runs.
- Release gates check the product before a build is shipped.
- Approvals hold anything that leaves your computer, and undo reverses what can be reversed.
- Signing proves a build and an update came from the maker.
- The data-flow audit checks what the app sends out against the code.
A model proposes, the rules check, you decide. No layer relies on the model being right.
Validators and never-allowed rules#
Before an action runs, the agent security rules look at it. Some actions are never allowed, whatever you approve: entering passwords or one-time codes, card numbers, changing operating-system privacy settings, granting system permissions and solving captchas. The list is data in config/agent-security/never-allowed.json.
Banking, health and government sites are treated as sensitive origins, with stricter handling. The list is config/agent-security/sensitive-origins.json. The protection levels you choose map onto these checks and add no new way to enforce anything (config/agent-security/presets.json).
Text that Growth OX writes for you is checked as well. A build check reads the site and listing text for overclaims and filler phrases and fails the build when it finds one (scripts/gates/overclaim-copy.mjs).
Release gates#
A security release runs every deterministic check: type checks, lint, the full test run, secret scanning, schema checks, the privacy gates and more. The only saving allowed is reusing a check whose input is identical to a recorded pass. Reviews by a person or a model are never replaced. The list is config/gates/release-tiers.json; the gate scripts are in scripts/gates.
The trust centre lists the gate scripts that exist in this version.
Approvals and undo#
Sends, posts, publishing, orders and spending arrive as approval cards. You read what would happen, then approve, edit or refuse. Under the most permissive preset, repeated sends of one kind stay automatic only until a daily cap is reached; after that you are asked again (config/approvals/daily-send-caps.json).
After an approved action runs, undo is available for a short time on the card and later from History. Some things cannot be taken back, such as a sent e-mail, and the card says so before you approve. See Approvals.
Signing#
Release builds are signed and checked on each platform: Developer ID signing and notarization on macOS, code signing on Windows and detached signatures on Linux. The release script refuses to run when an identity is not set, so a build is never silently unsigned. The files are config/release/macos-signing.json, config/release/windows-signing.json and config/release/linux-signing.json.
The app trusts the public keys in config/release/trust.json. A missing key means the app refuses the matching update or licence. It fails closed.
The data-flow audit#
Every feature that sends something out is listed in config/privacy/data-flow.json with what it sends, to whom and when. A test enforces Only this computer mode against that file, so a feature cannot start talking to the network unnoticed. The public page What leaves your computer is generated from the same file.
Questions#
- Does Growth OX hold certifications?
- This page makes no certification claim. It describes checks that run in the repository and names the file each one rests on.
- What if the model is wrong?
- Rules reject unsafe actions, you approve anything that leaves your computer, and undo reverses what can be reversed. The model is never the last check.
- Where can I see the current state of signing?
- On the trust centre. It reads the same configuration files and shows which identities are set.