Security

Sois runs the day-to-day operations of real businesses, and increasingly their AI agents act inside it on their behalf. That raises the bar: the platform has to be safe enough for a finance team, a customer-records system, and an autonomous agent to share the same workspace without anyone losing sleep. This page is the short version of how we approach that, written for the people who have to sign off on it.

It complements the Data protection page, which covers the GDPR controller/processor model and data-rights requests.

Isolation by design

Sois is multi-tenant, but tenants do not share a data store.

  • Every network gets its own database and its own object storage. Tenants within a network are isolated from one another, and networks are isolated from each other.
  • Isolation is enforced at the data layer, not just in application code — a request resolved for one tenant cannot read another tenant's rows or files, even in error paths.
  • The same boundary applies to agents. An agent operating in a workspace is scoped to that workspace and to the permissions of the identity it runs as.

The practical effect: there is no shared bucket, no shared table, and no "god query" that spans customers.

Encryption

  • In transit: everything is TLS (1.2+). HTTP is redirected to HTTPS, HSTS is set, and traffic to the platform terminates behind a managed edge.
  • At rest: databases and object storage are encrypted at rest with AES-256, using managed keys. This covers your business records, uploaded documents, and generated files alike.
  • Secrets and credentials — OAuth tokens, connected-account API keys, integration secrets — are encrypted in the database with application-layer encryption on top of at-rest encryption, and are never written to logs.

File storage

Files you and your agents create — documents, attachments, exports, reports — live in object storage that is:

  • Encrypted at rest and scoped per network/tenant, with no publicly listable buckets.
  • Served through short-lived, signed URLs, not permanent public links. Access is checked against the requesting identity's permissions before a link is issued.
  • Soft-deleted, not silently destroyed for user-facing records, so an accidental delete is recoverable rather than catastrophic.

Access control and authentication

  • Role- and permission-based access, applied least-privilege. A user — or an agent — only sees and does what their role allows, scoped to their workspace.
  • Agent actions are permission-gated at execution time. An agent cannot perform an action the underlying identity isn't allowed to perform; denials are graceful, not bypassable.
  • Authentication supports modern factors including passkeys / WebAuthn, with session controls and the ability for operators to enforce policy across their network.

Infrastructure and resilience

  • Hosted on major cloud infrastructure in private networking, with a managed edge providing TLS, WAF, and DDoS protection in front of the application.
  • Automated, encrypted backups with the ability to restore, plus point-in-time recovery on the primary data stores.
  • Deploys are zero-downtime and reversible — a release can be rolled back without a maintenance window.
  • Security-relevant events are logged for audit, and access to production is restricted and monitored.

What we deliberately don't do

  • We don't sit in your payment flow. Card processing runs on Stripe; Sois never stores raw card numbers and is not the holder of payment liability.
  • We don't sell or mine your data. Your business data is yours; we process it to run the service, not to build a product on top of it.
  • We don't expose the agent to anything the user can't already reach. The agent is a participant in your workspace, bound by the same permissions — not a privileged backdoor.

Compliance and disclosure

Our practices are built around the controls above, and formal attestations (such as SOC 2 and ISO 27001) are on our roadmap as we scale — if you have specific compliance, residency, or contractual requirements, we're happy to walk through them with your team.

If you believe you've found a security issue, please contact us at [email protected]. We welcome responsible disclosure and will work with you on it.