Risks & Guardrails

These tools are not designed to roam freely around your computer. They are designed to operate inside boundaries.

Sandboxing

The agent works inside a limited environment — usually just the project or folder you gave it. Not your whole computer.

Explicit approvals

It stops and asks before doing anything consequential. You stay in control of what actually happens.

You choose the mode

Configure how autonomous you want it to be. More locked-down for sensitive work, more freedom for routine tasks.

Best Practices

Treat it like cloud storage

Do not paste or upload anything you would not be comfortable storing in Box, Google Drive, or another shared cloud service.

Use least privilege

Do not give it access to email, your bank, payroll, calendars, or other accounts unless that access is truly necessary.

Keep secrets out

Avoid passwords, API keys, private tokens, SSNs, student records, medical data, and unpublished sensitive documents.

Start in a fenced sandbox

Give it one folder, one repo, or one task first, not your whole machine or all your accounts.

Review before it acts

Let it draft, summarize, or prepare actions first; approve important actions yourself.

Trust, but verify

Safety controls reduce risk, but they do not make the output correct.

8 / 12