Security · Updated Nov 2026
How Airbox keeps your agent contained.
An agent that can run commands will, sooner or later, do something you didn't mean it to: follow an instruction hidden in a web page, delete the wrong folder, or send a key somewhere it shouldn't. Airbox can't make an agent perfect. It limits what a mistake can reach. Here's what every box does, what it can't protect you from, and what we don't have yet.
The risks
What goes wrong when an agent runs on your laptop.
It reads things written by strangers
Web pages, emails, issues, attachments and chat messages can all carry instructions. That's prompt injection, and it doesn't need anyone to message your agent directly.
It can reach everything you can
On your laptop, that's your files, SSH keys, browser sessions, password manager and every API key in your shell. A confused agent can read or send any of it.
Its mistakes are hard to undo
An rm -rf in the wrong directory or a botched migration on your own machine has no undo button, unless you set one up beforehand.
What every box does
Six controls, on from the first minute.
- Isolation
-
A box of its own
The agent lives on a dedicated machine that holds nothing else of yours: no personal files, no browser sessions, no other accounts. It's separate from other customers' boxes, and the worst case stays inside it.
- Keys
-
API keys never touch the disk
Your model API keys are stored encrypted and added to requests at the network edge. They're never written to the box, so an agent that's been talked into dumping its environment has nothing to dump.
- Network
-
An outbound allowlist, on by default
The box can only reach the destinations it needs, like your model provider and package registries. That limits where an injected instruction can send your data. You can add destinations when a job needs them.
- Privileges
-
The agent runs without root
Agents run as an ordinary user, so a bad command can't rewrite the system, install a rootkit or switch off the controls above.
- Recovery
-
Snapshots, with a one-command restore
Snapshots run automatically. If the agent wrecks something on the box,
airbox restorerolls it back in seconds. - Supply chain
-
One company, every layer
Boxes run on hardware we own in a Miami data center. There's no hyperscaler or reseller underneath, so the list of companies that touch your box is us.
What a box can't do
Where containment stops.
-
It can't limit the permissions you hand over
Give an agent a GitHub token with write access and it can push to every repo that token can. Use fine-grained, single-repo tokens and read-only keys wherever you can.
-
It can't undo actions outside the box
A snapshot restores the box, not the world. Emails sent, messages posted, commits pushed and API calls made are already out there.
-
It can't stop misuse of allowed destinations
If an agent can reach a service, an injected instruction can try to use that service too. Keep the allowlist short, and keep tool-enabled agents on a strong model, since smaller models are easier to steer.
What we don't have yet
Honest about the gaps.
- No certifications yet. Airbox has no SOC 2, ISO 27001 or HIPAA certification today, and we don't sign BAAs. If you need those now, use a provider that has them.
- One region. Every box is in Miami. There's no EU data residency yet.
- Early access. We're racking boxes in small batches, and the
airboxCLI ships with the first batch. - No team features yet. SSO, roles and audit logs aren't built yet.
How we handle your account data, what we collect and how long we keep it are in our Privacy Policy.
Hardening guides
Running an agent on your own server?
The same ideas apply anywhere. These guides walk through each agent's own security settings, step by step.
How to run OpenClaw safely
Gateway on loopback, DM pairing, the tool sandbox, tool restrictions, secrets on disk and the security audit, from OpenClaw's docs.
Hermes on your own VPS
A non-root gateway service, user allowlists, the Docker terminal backend and key file permissions, from the Hermes docs.
What a cheap VPS leaves to you
The seven setup jobs between a fresh Linux server and one that's safe to leave an agent on.
FAQ
Security questions.
Is it safe to run Claude Code with --dangerously-skip-permissions on Airbox?
It's the situation a box is built for: the agent can only reach what's on the box and the destinations on its allowlist, and a snapshot undoes damage to the box. Keep the tokens you give it narrow, because a box can't stop an agent from using a GitHub token with write access exactly as that token allows.
Can Airbox staff see what's on my box?
We don't read the files, code, prompts or model output on your boxes. Staff access a box only when you ask us to, to investigate abuse or a security incident, or when the law requires it.
Where are my API keys stored?
Your model API keys are stored encrypted and injected at the network edge. They're never written to your box's disk, so an agent on the box can't read them.
Is Airbox SOC 2 compliant?
Not yet. Airbox is in early access and has no SOC 2, ISO 27001 or HIPAA certification today. If you need those now, use a provider that has them.
Where does my data live?
Boxes run on hardware we own and operate in a Miami data center. There's no hyperscaler or reseller underneath. Box disks and snapshots are erased within 30 days of deleting a box.
How do I report a security issue?
Email [email protected] with what you found and how to reproduce it. We'll confirm we got it, keep you updated while we fix it, and credit you if you'd like.