Security
What we keep, what we never keep, and who sees what.
Written for the person at your firm who has to answer for it. Everything here describes the system as it runs today. Where something is planned rather than done, it says so.
The one-line version
Chat messages, AI answers and the values we catch in them are never stored. Uploaded files are the one exception, described below. Everything in transit is encrypted with TLS, and the AI model receives placeholders, never the client values your firm's rules swap out. Anything your firm set to Flag is sent as written and reported to the owner. The only record we keep of a message is who sent it, when, and what kind of thing was caught.
The gate is updated continuously: new rules and identifier formats ship to every firm, and every update must pass our full automated attack suite before it goes live.
The one exception is files a staff member uploads to use with the AI connector in their own Claude or ChatGPT. The gate runs over the file as it arrives: passwords, keys and privileged material are refused and nothing is kept. Otherwise we keep two things, the protected text (placeholders for what your rules protect; anything set to Flag stays as written) and the caught values, both encrypted with AES-256-GCM and tied to that person's account, for 7 days or until they delete the file. The original file is never stored. The AI reads only the protected text; when it finishes a document, we fill in the values and keep the result, also encrypted, for 7 days, viewable only by that person when signed in. Encrypted copies of files and documents in our daily backups are gone within 14 days after that.
How a message is handled
| Step | What happens |
|---|---|
| 1. Check | The message is read against your industry template. Identifiers with checksums (cards, routing numbers, IBANs, VINs) are only counted when the checksum holds; SSNs must have a valid structure, and account numbers must sit next to a word like "account". |
| 2. Swap | Caught values become numbered placeholders such as [SSN_1]. Passwords, keys and privileged material are stopped outright; nothing is sent. |
| 3. Ask | Only the placeholder version goes to the AI model. |
| 4. Restore | The swap table travels with the conversation sealed (AES-256-GCM, bound to the account). Our server opens it for one request, puts the real values back into the answer, and never writes it to disk or to a log. |
| 5. Record | The owner's log gets a line: who, when, the outcome and the names of the rules that fired. Never the value. |
Controls in place
| Area | Control | Status |
|---|---|---|
| Encryption in transit | HTTPS only, with HSTS. Session cookies are HttpOnly, Secure and SameSite=Strict. | In place |
| Encryption at rest | Staff names and emails, and the activity trail, are encrypted in the database with AES-256-GCM; sign-in finds an account by a keyed hash of the email. Daily backups are encrypted whole. Keys live outside the data and web folders, readable by the service account only. | In place |
| Sealed swap table | The table that turns placeholders back into values leaves the server only sealed with AES-256-GCM and bound to one account. Browsers, the extension and API callers hold ciphertext they cannot read, alter or reuse for another account. | In place |
| Passwords | scrypt with a per-user salt. Sessions are random 256-bit tokens; only their hashes are stored. | In place |
| Firm isolation | Every query is scoped to the signed-in user's firm. An automated attack suite, run before every release, checks one firm cannot read or change another's data. | In place |
| Access roles | Staff and owners are separate roles; owner screens are refused to staff by the server, not only hidden. An IT admin gets the setup screens only; a password link an IT admin issues for a staff member is emailed to that person and never shown to the IT admin, so an IT admin can't sign in as a colleague. SOPHIA XT support can open a firm's setup screens (staff, rules, client list, topics and the API key list) for up to 60 minutes at a time, for a done-for-you setup or when the firm asks; every action is written to the firm's Activity as "SOPHIA XT setup (on behalf)", and it never reaches billing, the log, alerts or message text. | In place |
| Request safety | Strict content security policy, same-origin checks on every change, request size limits, rate limits on sign-in, sign-up and chat. | In place |
| Audit trail | Sign-ins, failed sign-ins, staff changes, setting changes and billing changes are recorded. Owners can read their firm's trail. | In place |
| Error monitoring | Server and browser errors are scrubbed before they are recorded: card numbers, SSNs, ITINs, passwords, keys, bank and routing numbers, EINs, medical record and health plan numbers, emails, phone numbers, dates of birth, and detected names and addresses. New errors alert us immediately. | In place |
| Backups | A consistent copy of the database is made daily; fourteen days are kept, readable by the service account only. | In place |
| Off-site backups | Copies stored with a second provider. | Planned |
| Two-factor sign-in | A second factor for owners and for our own admin access. | Planned |
| Names, places, dates | A firm's client list is matched exactly and stored only as keyed hashes. Other names, addresses and dates are detected from context, a best effort each firm can set to off, flag or protect. | In place |
| Independent testing | A third-party penetration test. | Planned |
Data we store, and for how long
| Data | Kept |
|---|---|
| Chat messages and AI answers | Never stored |
| Uploaded files for the AI connector: protected text and sealed values (never the original file) | 7 days, or until the person deletes them. Encrypted copies in our daily backups are gone within 14 days after that. |
| Documents finished through the connector | 7 days, encrypted, visible only to the person who made them. Encrypted copies in our daily backups are gone within 14 days after that. |
| The owner's log (who, when, what kind of thing was caught) | 400 days |
| Client list | Only as one-way keyed hashes, until the owner replaces it |
| Usage counts (messages and AI tokens per month) and what hosted AI cost each month | For billing, while the account exists |
| Activity trail | 400 days |
| Topics the owner sets up (name and phrases, encrypted) | Until the owner removes them. A topic the owner offers to other firms is reviewed by SOPHIA XT and, if approved, its name and phrases are shared with firms in the same industry, without the firm's name |
| Password reset requests (who asked, when) | 400 days after they are answered |
| Error records (scrubbed) | 90 days after last seen |
| Accounts: names, work emails, password hashes | While the account exists |
| The network address of each signed-in computer, to enforce one computer per person | 30 days |
| Revoked device sign-in records (extension and Windows app keys that were signed out or disconnected) | Deleted 30 days after they were revoked |
| Daily backups | 14 days |
| Your firm's own AI key, if the owner connects one | Until the owner removes it. Encrypted with AES-256-GCM, tied to the firm, and never shown again, only its last four characters |
| Card details | Never; Stripe holds them |
Who processes what
| Service | Purpose | What it receives |
|---|---|---|
| Hostinger | Hosting | The application and its database |
| Stripe | Billing | The owner's name, email, billing address and card, the number of seats, and what the firm's hosted AI use cost, in cents. Never message text or a client value. |
| An AI routing service, and the model maker your firm picks: Anthropic, OpenAI or Inception | Writing answers in the chat on Safe-Desk with the AI inside Safe-Desk, once Claude and ChatGPT models are offered there (today Mercury answers, through Inception, below). The routing service passes placeholder text to the model your firm chooses. If that model doesn't answer, another model your firm allows may answer instead, marked "backup"; a maker your firm didn't allow is never used | Message text and any files attached to it, with caught values already replaced by placeholders. Values your firm set to Flag are sent as written. Handled under SOPHIA XT's account |
| The AI provider your firm connects: Anthropic or OpenAI | Writing answers in the chat on Safe-Desk, only when the owner pastes the firm's own key | Message text and any files attached to it, with caught values already replaced by placeholders. Values your firm set to Flag are sent as written. The provider handles it under your firm's own account and terms |
| Inception (Mercury) | Only when Safe-Desk's own model is used: chat answers for a firm that has not connected a provider, and suggesting phrases when an owner asks for them on a topic | Message text and any attached files, or a topic name, with caught values already replaced by placeholders. Values your firm set to Flag are sent as written |
| AI chat sites (ChatGPT, Claude, Gemini, Copilot, Perplexity, DeepSeek), through the extension | Only when your staff use them | The placeholder version of what staff paste or send. The extension shows the answer with real values in its own panel; the chat page keeps the placeholders |
| ChatGPT or Claude, through the AI connector | Only when a staff member connects it | The placeholder version of files that person uploaded, their titles and the kinds of data replaced; never a value |
| The push service of the owner's browser maker (for example Google for Chrome, Apple for Safari) | Phone and browser alerts, if the owner turns them on | The alert line, encrypted to the owner's device so the push service can't read it. Never a value |
| ntfy.sh | Phone alerts, if the owner turns them on. Also our own alerts when someone asks for a password reset or the server records an error | The firm's name, the staff member's name and the names of rules that fired. Billing notices such as a failed payment or the monthly AI spending limit, on the owner's alert. For a password reset request, the person's role, with the firm's name on the owner's alert and only a firm number on ours. Our own error alerts carry the error message, scrubbed of client data, and the page it happened on. Never a value |
SOC 2
Safe-Desk has not been audited for SOC 2. We are building to the SOC 2 Trust Services Criteria for security, availability and confidentiality, and the controls above are the first part of that work. The remaining steps are written policies, the planned controls above, a compliance monitoring platform, and then an audit by an independent CPA firm. When a report exists we will say so here, and share it under NDA.
It also does not make your firm compliant on its own. It helps you meet confidentiality obligations by keeping client data out of AI tools and keeping a record.
Report a problem
Found a security issue? Email contactus@sophiaxt.com with "Security" in the subject. We read every report and will not take action against good-faith research.