Home / Where your data lives
Where your data lives
This page describes the service as of 17 November 2026. It has not been audited by an outside party. Questions: [email protected].
Last updated: 17 November 2026
Contents
The path your data takes
The hardware is ours and is in the Philippines. If you need to know more about where it is, ask us at [email protected] before you send us any data.
Who can log in
- Your people. Sign-up is by invitation. You invite your own team and give each person a role, and the role limits what that person can see and do.
- SunderSystems. One person, the founder, has administrative access to the hardware and the platform. What staff do in the admin area is recorded in a staff audit log.
- No one else. Other clients have no access to your workspace, and each hosted desktop runs in its own container. The containers share one machine; they are not separate virtual machines.
What stays on our hardware
Your workspace data and your hosted desktop run on our own hardware in the Philippines. When a request goes to an open-weight model that runs on our own GPU, or to your own model once it is trained, the prompt and the answer are processed here and are not sent to any AI provider.
Each hosted desktop is a Docker container on one shared machine. It is not a virtual machine, and it is not a machine of its own for each client. A desktop can open public websites, as a person’s browser can. A machine used by one client only is planned and not available yet.
Traffic between your browser and our hardware passes through Cloudflare, which carries it to us through a tunnel. The other outside services we use are on the sub-processors page.
What leaves our hardware
AI requests go to open-weight models only: a model on our own hardware, or an open-weight model through OpenRouter. Each request sent through OpenRouter, with its prompt and context, goes to OpenRouter and to the provider OpenRouter routes it to, under that provider’s terms. OpenRouter is told to use only providers that do not collect or train on prompts, that take every parameter we send, and that charge no more than $0.30 in and $1.20 out per million tokens. We have not limited the countries those providers may be in. If you add your own provider key, your requests go to that provider under your own agreement with it. The app shows the price of each model your workspace can use.
We do not offer closed models. The platform refuses them.
For developer API calls we store usage numbers only: the model, the tokens, the cost and the time taken. We do not store the prompt or the reply.
The approval step
Each workspace has an autonomy setting: Draft only, Ask me first, or Act and tell me. With the first two, sending a message, posting something or another write action waits for a person to approve it. With Act and tell me, routine tasks go out without waiting. The setting is yours to choose; we do not change it for you. You can also watch the hosted desktop live from the web dashboard while the agent works.
Actions the AI takes, and the approvals, are recorded in a log you can review. Each record keeps the tool, its inputs with known secret formats (keys, tokens, passwords) hidden, the result and who approved it. Free text may still contain sensitive words, so do not paste secrets into prompts. It is kept for 90 days, then deleted automatically, and workspace owners and admins can export it before then.
Account protections
Available on the platform today:
- Passwords stored as hashes, never as the password itself.
- Two-step sign-in with an authenticator app.
- A devices and sessions page, where you can see where you are signed in.
- Team invitations with roles.
- A staff audit log.
Email confirmation and password reset by email are built and are sent through Mailgun from mg.sundersystems.com once that sending domain is verified. Our mail log keeps no address and no content, and is deleted after 90 days. If an email does not arrive, write to us.
Training
Training needs your consent. It happens only if an owner of the client’s workspace turns it on in settings, and we record that consent. It runs on our own hardware, and the model is used for that client only. A client’s data is never used to train a model for another client. Your data trains only your own model.
We test each client model against a gate before it goes live. A model that does not pass is not used. Chat histories you import from other tools are stored read-only and are never used for training. Text written by Claude or another closed model is never used for training. Messages your staff label on the labelling page stay in your workspace, are never used for training, and are erased by “Delete model and data”. The shared default models are untuned open-weight models.
Encryption and backups
Traffic to our sites and the app is encrypted in transit with TLS, through Cloudflare. The platform database is backed up every night and the nightly backup is encrypted. We keep 14 daily and up to 8 weekly copies, so a backup ages out within about ten weeks. The backups are kept on our own local disks and on a USB copy. There is no cloud copy, so they do not protect against everything, such as losing the whole site. A restore test of the database and of one desktop passed on 12 November 2026.
We do not claim that data stored on the hardware is encrypted at rest. Desktop home folders, which can include browser sign-ins, are backed up every night. The copies are encrypted and kept on our own storage, 14 daily and up to 8 weekly, and a deleted desktop’s copies age out in about ten weeks. For security we log the host names (not page addresses or content) that your AI desktop connects to, and keep that log for up to 30 days.
No certifications yet
We hold no security certifications today. Preparing for a SOC 2 audit and registering with the Philippines National Privacy Commission are both planned. Neither has been done. We will not claim a certification until we hold it.
If your work is regulated, check your own rules before you send us data.
One person on call
SunderSystems is run by one person today, the founder, who watches the service and answers problems personally. There is no uptime promise and no public status page yet. Both are planned.
If something goes wrong
If a personal data breach occurs, we investigate and contain it. Our target is to tell affected business clients within 72 hours of becoming aware of it, as our data processing addendum says. Where the law requires it, we notify the Philippines National Privacy Commission and the people affected within 72 hours.
How to report a problem
If you find a security problem, email [email protected] with the subject “Security”. Give us reasonable time to fix it before you make it public, and do not access other people’s data or disrupt the service. We will acknowledge your report and keep you informed. Our security.txt file has the same contact.