Public beta disclosure
Risk disclosure
The most important fact about Breezable is simple: its AI helpers are not locked inside a safe area.
The one thing you must understand
Breezable runs AI helpers on your computer. An AI helper is a program that reads instructions and can work with files or run computer commands. Today, those helpers are not locked in a sandbox.
A sandbox is a locked area that limits what a program can reach. Breezable does not currently place its helpers in one. When you pick a project folder, that folder is only where the helper starts. It is not a fence.
A helper can reach anything your Windows account can reach, including:
- documents, downloads, photos, and other projects;
- saved sign-in files and credentials;
- command-window history and results;
- running programs; and
- internet or local-network services your computer can reach.
What “no sandbox” means
The helpers run like ordinary programs under your Windows account. They can read, change, or delete files outside the folder you selected. They can run commands. They can use credentials already available on your computer. They can send relevant instructions, code, and context to AI companies such as Anthropic or OpenAI.
Breezable starts helpers with their normal permission questions turned off. You should not expect a helper to ask before every file change or command.
Why this is risky even if nobody targets you
AI helpers read text from your instructions, files, project notes, webpages, and command results. Some text can contain instructions aimed at the AI instead of you. This is called prompt injection: text tries to trick the helper into ignoring your request and doing something else.
If a helper is tricked or simply makes a mistake, Breezable has no sandbox to contain what happens next.
What could go wrong
A helper could make a mistake or be tricked into:
- deleting or overwriting files;
- exposing passwords, service keys, or private information in a request, saved record, command, or problem report;
- changing a project in a way that breaks it;
- running a command you would not have approved;
- using your signed-in Claude, Codex/OpenAI, GitHub, or other credentials; or
- sending sensitive code or audio to another company when the selected feature does that.
How to reduce the risk
These steps do not make Breezable safe. They reduce how much is at stake.
- Use a separate Windows account without personal files or important sign-ins.
- Back up important files before letting helpers work.
- Only open projects you trust. Treat downloaded projects, copied webpages, and files from strangers as risky.
- Do not use secrets, customer information, regulated information, or private company code unless you have permission and understand the rules of every AI company involved.
- Review all changes before running or sharing them. AI-generated work can be wrong or insecure.
- Turn off raw command-window recording if you do not want the text you type and receive saved locally.
- Avoid Cloud dictation for sensitive speech. The Cloud option sends recorded audio to OpenAI.
- Leave remote access features off unless you specifically understand and want them.
Only use Breezable on a computer, Windows account, and projects where you accept this access. A separate account or disposable computer is the safer way to explore the beta.