Security and AI privacy

This page answers one question in plain words: what happens to a message when it arrives at xmail, and who else sees it. It describes what the service does today, 20 October 2026, not what it may do later.

The short answer

• A free account has no AI features, and nothing from its mail is sent to any model, ours or anyone else's.
• With a paid name, every incoming message is read by our own model, running on our own machines, to say what the message is and whether it looks risky.
• Three features send text to an outside company, and only when you ask for them: translating a message, drafting a reply, and the advisor that reads a whole conversation.
• xmail is not end-to-end encrypted. We can read the mail in your mailbox, because the service has to in order to deliver, index and analyse it. Anyone who tells you their AI mailbox is zero-knowledge is describing something else.

What runs where

Classification — the part that happens to every message — runs on our own GPUs, in our own racks, on an open-weights model (Qwen 2.5). Nothing about that step leaves our network. The quick filters before it are ordinary rules and statistics, and they run on the same servers.

In the thirty days to 20 October 2026 the service made 13 627 model calls. 13 564 of them ran on our own hardware; 56 went to Anthropic (message translation and reply drafts) and 7 to OpenAI (the thread advisor). That ratio is what the architecture produces: the automatic work stays in-house, and only a feature you press a button for goes out.

What goes to an outside company, and when

Both companies state that data sent through their APIs is not used to train their models, and that they delete it within 30 days unless their own safety systems flag it: OpenAI and Anthropic. We use ordinary API accounts, so those default terms are the ones that apply; we do not claim a special zero-retention agreement, because we have none.

Models from other providers are not in the list that may receive message text. That is enforced in the code by an allow-list, not by a promise.

Training

Outside providers do not train on what we send them, as quoted above. We do improve our own classifier on our own servers, using stored decisions that include the first part of a message. If that matters to you, it is the honest picture: your mail never becomes training data for OpenAI or Anthropic, and it does contribute to the model that sorts mail inside xmail.

What is stored, and for how long

• The mail itself stays in your mailbox for as long as the account exists. That is what a mailbox is.
• A cached copy of recently read bodies lives for seven days to keep the interface fast, then disappears by itself.
• What the AI concluded — the type of the message, the risk score, extracted dates and amounts — is kept with the message.
• Deleted mail leaves the Trash after 30 days.
• There is no other automatic deletion by age today. We would rather write that down than imply a retention policy we do not run.

Deleting your account

Closing the account removes the mailbox itself, the messages in it, and the data built from them across the database — including the entries in the search index that holds the sender, the subject and the summary of each message. One thing takes a little longer, and we say so rather than round it up: the seven-day cache of recently read bodies expires on its own instead of being wiped on the spot. A record that the account existed, the payments and the points ledger remain, because tax and billing rules require them.

There is no self-service export of your data yet. Until there is, mail can be copied out over IMAP with any mail client — see the mail-app settings.

The ordinary security parts

• Every connection is encrypted: HTTPS for the site, TLS for IMAP and SMTP. Unencrypted mail protocols are refused.
• Passwords are stored hashed, never in readable form.
• Incoming mail is checked for SPF, DKIM and DMARC alignment, and the result is shown to you, not hidden.
• Sending limits apply to new accounts, so a stolen account cannot be used for a mass mailing.
• We do not run advertising or third-party tracking scripts on these pages.

What we do not claim

No end-to-end encryption for mail. No zero-knowledge storage. No security certification — we have not been audited against SOC 2, ISO 27001 or HIPAA, and we will not pretend otherwise. If a page somewhere states one of these about xmail, it is wrong and we want to know.

Reporting a problem

Write to [email protected]with the word “security” in the subject. Tell us what you found and how to reproduce it; we answer. Please do not test against other people's accounts.

Related: the privacy policy, the terms, and the FAQ.