A phone sandbox for AI agents: let the agent use its own Android phone, not yours.
Why not your own phone
An agent that can tap anything can open anything: your messages, your email, your bank, your photos. Most agent mistakes are small, like a wrong tap or a misread screen, but on your personal phone a small mistake can send a message or approve a payment. Prompt injection makes it worse: text the agent reads on one screen can try to steer what it does on the next. A separate phone limits what any mistake can reach to what you put there.
What a good phone sandbox gives you
| Control | Why it matters | On Phonebox |
|---|---|---|
| Only the apps you choose | Nothing else to open by mistake | A fresh phone has no accounts; install apps from an APK, a link or the app library |
| A key limited to one phone | A leaked or misused key reaches nothing else | Keys limited to certain phones, with Agent, Read-only or Admin access and an expiry |
| No delete by default | The agent can't destroy phones or their data | Agent keys create, use and park phones but can't delete them |
| A person for sign-ins | Your password never passes through the agent | A live link opens the phone in any browser so you sign in yourself |
| A record of what happened | You can see and replay what the agent did | An activity log of every call, with the screen kept on failures, and screen recordings |
| A spending cap | A stuck loop can't run up a bill | A monthly spending limit, and each session's max_duration, up to 3 hours |
| A clean slate | No leftovers between tasks or users | Reset wipes the phone back to clean from the console, API, SDK or CLI |
Set one up
- Create a phone for the task in the console, or connect your agent with
phonebox setupand let it create one. - Install only what the task needs, and sign in to those apps yourself through a live link.
- Create a key limited to that phone under API keys, with an expiry, and give the agent that key rather than your own.
- Set a spending limit in Settings, and record the first runs so you can watch what the agent does.
- Reset the phone when the job ends, or park it to keep its apps and sign-ins for next time. A parked phone costs nothing.
What a sandbox doesn't fix
A sandbox limits what a mistake can reach; it doesn't stop the agent from misusing the accounts you sign in to on that phone. Use test or low-risk accounts where you can, avoid payment apps unless the task needs them, and read each app's terms: some forbid automated use. Phonebox phones have no SIM, so the agent can't send texts or make calls from them.
Questions
How do I give an AI agent a phone safely?
Give it a separate cloud Android phone instead of your own, with only the apps the task needs. On Phonebox, give the agent a key limited to that phone, sign in to apps yourself through a live link, set a monthly spending limit, review the activity log and recordings, and reset the phone when the job is done.
What is a phone sandbox for AI agents?
A separate Android phone, usually in the cloud, that an AI agent controls instead of your personal device. It holds only the apps and accounts you choose, so a wrong tap or a prompt injection can only reach what's on that phone.
Can an AI agent delete my phones or data on Phonebox?
Not with an Agent key, the default for agents: it can create, use and park phones but not delete them. Deleting needs an Admin key, which the console only creates after you confirm it. Read-only keys can only look.
Get started with Phonebox. New accounts start with a free 7-day trial, card required.
The Phonebox team