Security

Where your data lives

• Database, sign in and uploaded files: Supabase, Frankfurt (eu-central-1).
• Server functions: Vercel, pinned to Frankfurt (fra1).
• AI, only when switched on: Anthropic. It processes in the US or globally. There is no EU region.
• Email: Resend, when email is configured on the deployment. It sends evidence requests, reminders and invitations.

Access and isolation

• People sign in with Supabase Auth.
• Every client record belongs to an organisation. Row level security is on for every table, and most reads and writes run as the signed in user, so the database applies the rules.
• Roles are firm admin, firm member and client company users, plus owner, reviewer, contributor, viewer and client reviewer on each project. The database enforces most of these rules.
• Some actions, such as deletions, membership changes, logging and file storage, use a privileged database key. Each checks the caller in application code first.
• Significant actions go to an audit log that only platform admins can read. We do not call it tamper proof.

How AI is used

Treat every AI output as a first draft for a person to check. AI drafted disclosures cannot be exported until someone ticks that they have reviewed them and take responsibility for them. Benchmarking drafts show the same acknowledgement but do not lock anything. Answers from the assistants have no sign off step.

• AI is off by default for every organisation. A project owner must also consent before that project's AI features run. A consultancy admin can switch AI off for the whole firm or for one client company. The strictest setting wins. When AI is off, nothing is sent: you get the rules engine result, or no answer from the assistants.
• Calls run on the server. Each route sends named fields only, and uploaded evidence files are never sent.
• Each firm has a monthly AI budget. Every call is logged with tokens, model and data posture, and the log holds no prompt or response text.
• Drafts you accept are stored in Kateri like any other content.
• Anthropic runs under its standard commercial API terms and does not train on inputs without express permission. Kateri calls none of the models Anthropic keeps content from for 30 days, so prompts and outputs are not retained by default. If Anthropic flags a session it may keep it for up to 2 years, or longer where the law requires (Anthropic's terms as of 20 October 2026).
• We have no zero data retention agreement with Anthropic.

Other controls

• Evidence uploads: 10 MB limit, an allowlist of file types, and the first bytes of each file are checked against the type it claims.
• Evidence download links expire after one hour.
• Email: one user can trigger 20 evidence, reminder and invitation emails an hour.
• Setting a password through a reset or invite link asks Supabase to sign out every other session.
• Signing out clears the Kateri data cached in the browser. Data stays on a device that never signs out.
• Traffic uses HTTPS. Supabase encrypts the database and storage at rest. Both are provider controls.

Not in place yet

• SOC 2 and ISO 27001: not held. We plan to pursue SOC 2 when a buyer asks. No date is set.
• Multi factor authentication: none yet. It is planned first for platform admins.
• Penetration test: planned, after multi factor authentication. No date is set.
• Zero data retention with Anthropic, and EU region inference: planned, no agreement signed.
• Deleting a person's account and erasing their data are done by hand by us.

Report a security problem

No security inbox is published yet. It will be listed here.