Security and control

How AROS keeps work safe

The controls below are enforced by the software today. What we do not claim is listed at the end.

Each company is kept apart

Every record belongs to one company, and the app checks which company a record belongs to before reading or changing it by its id.
People see their own company's data according to their role: viewer, operator or CEO.
The platform operator's console shows a company's name, plan and usage - not its leads, prospects, communications or knowledge - and every operator action is recorded in the company's own audit trail as well as the operator's.

Every external action an agent requests is checked against the agent's autonomy level and the action's risk class before it leaves AROS. By default, publishing, sending and spending wait for a person; CRITICAL actions always do.
Approvals route by risk: high-risk and critical actions need a CEO.
A kill switch stops autonomous external action, and it is re-checked at the moment of execution, so an earlier approval does not outlive it. Pausing an integration or an agent stops actions already approved for it.
Daily and hourly action budgets per integration hold work once a limit is reached instead of sending it; nothing approved is discarded. AI spend has a hard daily and monthly budget.

Agents produce text and structured data. They hold no credentials, and they can only request capabilities from a fixed catalog - they cannot invent an integration or call an arbitrary address.
Every model call carries a notice that fenced text is data, not instructions. An agent's task input - incoming messages, partner events, earlier agents' output - travels fenced, and injection-shaped text inside a fence is flagged in the audit trail.
The deployment's secrets are removed from everything sent to an AI provider. AROS trains no models on your data.
Who a message goes to comes from the record that started the work, never from what a model wrote.

Integration credentials and authenticator seeds are sealed with AES-256-GCM under a key kept outside the database, and are never shown again after they are saved.
Incoming webhooks are verified by signature before anything they carry is stored.
A payment is credited only after AROS re-checks the amount, currency and reference on the server - never because a browser says it was paid. No card details are stored.

Accounts are by invitation. Sign-in is by password, Google or a one-time code, and repeated failures are throttled per address and per network.
An unknown address and a wrong password get the same answer in the same time, so sign-in does not reveal who has an account.
Google sign-in tokens are verified on the server with Google's own library. One-time codes expire after ten minutes and five wrong attempts.
Suspending an account, resetting its password or signing out everywhere ends its existing sessions on their next request.

Every external action, approval decision and AI cost is recorded, with secrets redacted before the record is written. No screen edits or deletes an audit entry.
Requests to access or erase personal data are supported. Erasure pseudonymises the audit trail rather than deleting it, so the history of decisions survives without naming the person.

Before a change is released, real-browser simulations walk the whole app, a two-company walkthrough checks that neither can see the other, and the build is scanned for committed credentials and for production dependencies with critical security advisories.
A release reaches users only after the new version answers its own health check, including a check that the database schema matches it.

AROS holds no third-party security certification (such as ISO/IEC 27001 or SOC 2) today.
Companies share one hosted database; the separation between them is enforced by the application on every request, not by separate databases.
AROS is hosted by SnxwFairies on Google Cloud Run in the Mumbai region. There is no self-hosted edition.
The legal pages are drafts awaiting counsel and say so.

To report a security concern, or to ask for more detail before you buy, contact us. More answers are on the questions page.