Legal

Privacy Policy

Last updated: November 21, 2026

1. Information We Collect

SiftMail collects only the information necessary to provide email triage services. This section is the complete list of what SiftMail reads from your mailbox and what it keeps; Section 6 names everyone outside SiftMail that receives any of it.

  • Account data: Your email address, your OAuth provider (Google or Microsoft) and, if you connect an Outlook mailbox, the account and directory identifiers Microsoft issues for it (used only to recognise that account when you sign in again), your plan, and your settings, including the allow, block, VIP and trusted-sender lists and the saved searches you create. If you choose a send time for your daily digest, SiftMail keeps that hour and your browser's time zone name (for example, America/Los_Angeles), and uses them only to send the digest at that hour.
  • Email content: SiftMail reads a message's sender, subject and other headers, its Gmail labels, Gmail's short preview snippet, its body, and the names, types and sizes of its attachments. SiftMail never downloads attachment files. It also reads the names of your Gmail labels and how many messages and conversations your inbox holds. When you connect, it reads the 500 most recent messages in your inbox, and after that new inbox mail as it arrives, which it checks for every few minutes. Once the 12-month history fill is switched on for your account, SiftMail also reads, once, the sender, recipients (To and Cc), subject, date, reply and message-ID headers, unsubscribe headers, Gmail labels and Gmail's short preview snippet of up to 10,000 messages from the past 12 months, including archived mail and mail you sent but never spam, trash, drafts or chats. It asks Gmail for these headers only, so no message body is read for it, and it sends nothing to Anthropic or any other AI model. It keeps them as message records (below) so Money, Life admin, People, Ask, search and Rediscover's year summary can cover the past year instead of only your newest mail, and about once a month it checks again for messages it missed. Those records carry no score and are never used to decide whether a message is a threat. Until you finish or dismiss the one-time cleanup the dashboard offers you, opening the dashboard prepares a preview of it without waiting to be asked: it reads the headers and Gmail's snippet of up to 2,000 inbox messages in each of its three cleanup groups (promotions, bulk mail and old no-reply mail), and nothing is archived unless you confirm (Section 4). It reads other mail only for a feature you start or for one of three background checks. The features are: an inbox cleanup (Section 4), which reads the headers and Gmail's snippet of the inbox mail it covers (a threat cleanup covers all of your inbox and can pass what it reads to security scoring, Section 6); the Sift Inbox Index, which reads all of your mail, of any age, except sent mail, drafts, spam and trash; the storage screen, which uses only the sender, subject, date, size and labels of large messages with attachments; and Rediscover (Section 8), which reads messages that have attachments. SiftMail pages that list your inbox, such as the dashboard, read the headers and snippets of the messages they show. The background checks are: several times a day SiftMail re-reads, in full, the messages carrying its Sift/Quarantine label, re-checks them by its own rules without AI, and releases any that no longer look like a threat (Section 4); once a day it reads the labels and dates of the messages in each recent thread it marked as needing a reply, including the replies you sent, to see whether you have answered; and, only once recurring maintenance has been switched on for your account, each run lists your recent inbox messages and judges them from what SiftMail already stored about them (Section 4). SiftMail uses what it reads for: (a) security scoring (phishing, malware, spam and promotional classification); (b) judging whether a message needs a reply, is safe to archive, or was written by a person, which feeds your dashboard, your digest and recurring maintenance (Section 4); (c) the Sift Inbox Index, which builds a searchable, structured index of your mail (e.g., receipts, travel, subscriptions) so you can find buried information by natural language search; (d) the Exposure Index, an automated sensitivity scan described in Section 3 that runs on every incoming message and on every message the index reads; (e) cleanups and the storage screen (Section 4); and (f) the optional Rediscover deep scan (also offered as the one-time Decade Audit), described in Section 8. Email content is processed by automated systems to provide these features. SiftMail does not allow humans to read your email content except in the narrow, disclosed circumstances described in Section 7 (Google API Limited Use Disclosure).
  • Message records: For every message SiftMail scores, it keeps the sender, subject and date, all of the message's headers, its Gmail labels, Gmail's short preview snippet and the links or link domains found in it, together with the result: a label (such as safe, promotional, spam or phishing), a risk score, the reasons behind the score, the message's sender-authentication results, and whether it appears to need a reply, be safe to archive, or come from a person. The reasons can include the AI classifier's short written assessment of the message and details such as a link, domain or reply-to address it found suspicious. Messages read by the 12-month history fill (above) are kept as the same kind of record, holding only the headers listed there, the labels and the snippet, with no score, until you disconnect or delete your account like every other message record. On the Business plan, a check every six hours over the past week's message records also keeps a note of each sender whose reply-to address changed or whose links pointed at a lookalike of the sender's own domain, naming the sender and the addresses or domains involved.
  • Index entries: The Sift Inbox Index does not only read your mail, it stores what it extracts. For a message it recognises — a receipt, an order, a booking, a shipping notice, a statement, a subscription charge — SiftMail keeps the sender and sender name, the subject, the date, a Gmail snippet, a short title, the category it assigned, the amount and currency, the merchant, any location named, whether the message had attachments and their file types, and the structured details it read out of the message for that kind of mail. Those details differ by kind. For example: for a receipt, the order number, payment method, subtotal, tax, total and return-by date; for travel, the confirmation code, the airline or hotel, the airports, and the travel dates and times; for a shipping notice, the carrier, the tracking number and the estimated delivery; for a bank, payout or tax message, the institution, the last four digits of the account, the transaction type, the balance, the period and the tax year; for a contract or other document, its type, the other party and any sign-by date; for a subscription, the service, plan, billing cycle and renewal or cancellation date; for a restaurant booking, the cuisine, party size, time and delivery service; for an event, its name, venue, times and organiser; for an account notice, the service and whether action is needed; and for a shared photo, the number of images, the album service, the occasion and the people mentioned with it. The classifier can return details beyond these, and SiftMail keeps them too, apart from lists of the items bought, which it looks for and removes before storage. There is no separate purchase date or shipping cost; the message's own date stands in for when you bought something. This is how search for "what did I spend at that hardware store in March" works at all. SiftMail records the totals of a purchase and never what you bought — no product names, quantities or per-item prices are requested from the classifier or kept in your index. SiftMail does not store the message body itself; it stores these extracted fields and the Gmail message ID so you can open the original. For every message the classifier has read for the index — during an index run, or when the message arrived — including one it found nothing in, SiftMail also keeps that message's ID and the time it was read, and nothing else from it, so the index does not send the same message to the classifier again. Mail SiftMail recognises as legal or health-related is not sent to the classifier for the index at all (Section 6).
  • Exposure findings: When the sensitivity scan described in Section 3 flags something, SiftMail stores the finding — its type, category, severity, the sender, subject and date of the message it came from, any attachment file name involved, and a short excerpt of the surrounding text, masked as Section 3 describes. Where the sensitive information belongs to someone other than you (for example a colleague named in a forwarded document), the finding may record that person's name or role so you can see whose data you are holding.
  • Contact details: Names, email addresses, phone numbers, and company names that the AI classifier finds in your mail are kept in a per-user contact directory shown in your SiftMail dashboard. Mail from a sender in this directory is treated as mail from someone you correspond with, which keeps it out of recurring maintenance (Section 4).
  • Confirmed contact merges: The People screen can suggest that two of your contacts' addresses belong to the same person. If you confirm, SiftMail stores that pair of addresses, the date, and how the link was made, so the two show as one person on screen. Nothing is stored for a suggestion you have not confirmed, and you can undo a merge at any time (“Separate”). A merge only changes how the People screen groups addresses; it never makes an address count as a known contact for security scoring. Everything else on the People screen (how often you write to someone, threads waiting on a reply, things you said you would do) is worked out from the message records above each time you open it and is not stored separately.
  • Vouched senders: When you mark a flagged message as safe (Not a threat) and that message was properly authenticated (it passed DMARC and was signed by a domain matching the sender's address) with no strong threat signal, SiftMail records a vouch for that sender. This feature is switched on. A vouch stores only the sender's exact address, the domain that signed the message, that it came from a Not a threat correction, when it was made, how many times it has been used, and when you revoked it if you did. It stores no message content. A vouch can only lower the score of a later message from that exact address when it is signed by the same kind of aligned domain, and it lapses after 365 days. It never overrides an Allow, Block or VIP rule you set (your rule decides), and it never overrides a strong threat signal such as a lookalike domain or a failed signature. The Trust screen shows your vouches beside your rules, and you can revoke any of them there.
  • Not a reply: When you tap Not a reply on a message in the needs-reply list on Today, SiftMail stops listing that message as needing a reply. If SiftMail had already judged the sender to be automated (not a person), it also stores that sender's full address, the time you tapped, a running tally of later messages it applied to, and when you revoked it if you did, so later mail from that address is not listed as needing a reply. It stores no message content. It only affects the needs-reply list: it never changes a threat check, a label or anything in your mailbox. The Trust screen lists these senders, and you can revoke any of them there.
  • Screener (first-time senders): On Pro and Business, SiftMail can show mail from an exact From address it has not seen before in the mail it has checked for your account (not your whole mailbox history) in a Screener list on Today and in your digest, so you can Allow, VIP or quarantine that sender. A lookalike or DMARC-fail sender cannot be one-tap Allowed or VIP'd. For those accounts SiftMail stores a “first-time sender” marker alongside that message's existing score record. When SiftMail cannot check a sender's history, it does not mark the sender as new. Screener stores no new message content; Allow and VIP write to the same allow/VIP lists already described above, and quarantine uses the existing Sift/Quarantine label in Gmail or SiftMail Quarantine folder in Outlook. Free accounts do not get the Screener list, and SiftMail does not mark their senders as first-time.
  • One-tap digest links: Once one-tap digest actions are switched on (they are not yet), a flagged message in your digest can carry Release, Keep or Not a threat links. Each link is signed for your account, names the message by its Gmail ID, and works once, for 48 hours. Opening it only shows a confirm page; nothing changes until you press the button. When you use one, SiftMail stores a random one-time code for the link, your email address, the action and when you used it and when the link expires, so the link cannot be used twice. It stores no message content, and clears each code once its link has expired, or when you disconnect.
  • Password breach check: On Pro and Business, SiftMail tells you whether a password found in your mail appears in known breaches. It sends a k-anonymity range query to Have I Been Pwned's Pwned Passwords service (api.pwnedpasswords.com): only the first 5 characters of a one-way hash (the first 5 hex characters of the SHA-1) of that password, with padding so the reply does not reveal which hash was asked about, and compares the rest on its own servers. If the password as written ends in sentence punctuation, SiftMail also sends a second, separate range query for the punctuation-trimmed spelling — still only the first 5 hex characters of that spelling's SHA-1, and never more than two prefixes per detected password. Each detected password is checked at most once a day. The password, its full hash, your email address and the message are never sent. SiftMail keeps only whether the password was found and when it checked, and clears those two fields after 30 days. Free accounts do not get this check, and SiftMail does not send any hash prefix for them.
  • Scoring records: For each message SiftMail scores from now on, it keeps a small record, alongside the verdict, of how that verdict was reached: the version of the scoring rules, a fingerprint of their settings, the types of risk signal that fired and how much each counted, which AI model was used (only when AI was used), and any rule that lowered the score. It contains no message content: no subject, body, snippet or address beyond what the message record above already holds.
  • Tamper-evident action history: The record of actions described under Usage data below now notes who took each action (you or SiftMail) and links each entry to the one before it with a cryptographic fingerprint, so that an entry altered or removed after the fact can be detected. You can check your own history on the Trust screen. The fingerprints are computed from the entries themselves and add no mailbox content.
  • Money, Life admin and Ask: These screens (bills and subscriptions, spending, returns and refunds, tax forms, trips, insurance, legal deadlines, car, home and school, and asking questions of your mail) are worked out each time you open them from the sender, subject, date and Gmail snippet SiftMail already keeps in its message records (and, for a few of them, the To, Cc and unsubscribe headers already held there). They store no message bodies and nothing new of their own, apart from the three kinds of record below that you create yourself. They make no AI call: nothing on them is sent to Anthropic or to any other AI model, so legal mail is never sent to an AI model by them. Health-related mail gets no Health topic on the Life admin screen and is left out of search results and Ask answers unless the health topic is switched on, and it is currently switched off. Every other optional part of these features (model-written answers and drafts, sending, saving drafts into Gmail, Gmail labels for snoozes, calendar inserts, form filling and Slack, text or push summaries) is also currently switched off, and we will update this policy before any of them is turned on. On Pro and Business you can create a private calendar subscription link for your own calendar app: SiftMail stores only a one-way hash of that link's token and when you created it, never the token itself; your calendar provider (for example Google, Apple or Outlook) fetches, and keeps on its servers, due dates with the payee or insurer, the kind of date and the trip destination, taken from mail SiftMail already stores (about the last 120 days); amounts, account and case numbers, legal dates and mail SiftMail recognises as health-related are left out (that recognition is automated and can miss some health mail); anyone who has the link can see those dates; you can replace or turn the link off at once, which stops the old URL from working; Railway may record the path of that link, including the secret token in the URL (Section 5).
  • Saved facts: SiftMail can suggest details from your mail worth keeping, such as a policy or membership number, and you can choose to save them. This feature is switched on; nothing is saved unless you save it. Each saved fact's value is encrypted with its own key (AES-256-GCM, bound to your account, that record and that field, so it cannot be decrypted if copied elsewhere), and that per-fact key is itself stored only in encrypted form, wrapped by a master key. Alongside the encrypted value SiftMail stores the fact's type, the sender or merchant it came from, the message it came from, its last four characters (shown in the masked list) and a keyed fingerprint used to find it, which cannot be reversed to the value without a separate secret key. Every reveal, export and deletion is recorded in your action history by count only, never the value. The master key and the fingerprint key are held in SiftMail's server configuration on Railway, not in a separate key-management service. The encryption protects the saved copy only: the value also appears, unencrypted, in the subject or snippet of the message record it came from, and for a card or account ending the stored last four characters are the whole value (Section 5). You can reveal, export, delete one or delete all of your saved facts at any time.
  • Reply drafts: When you ask SiftMail to write a reply or a cancellation email, it stores the draft (who it is addressed to, its subject and text, the message it answers) with a fingerprint of its content and whether you approved it. Drafts are built from templates and the words you type; no AI model writes them. A draft cannot add recipients who are not already on the thread unless you typed them. Draft records are deleted 30 days after they are created, or as soon as you discard one. Nothing is sent without your explicit approval, and sending from SiftMail is currently switched off altogether: you copy the text or open the thread in Gmail and send it yourself.
  • Follow-ups and snoozes: When you snooze a message or ask to be reminded about it, SiftMail stores the message and thread it refers to, whether it is a snooze or a follow-up, when it is due and whether it is done. These are kept in SiftMail only: nothing in your Gmail changes. The record of a finished or cancelled one is deleted 90 days after its due time.
  • Deep-scan findings: What the optional Rediscover deep scan keeps is listed in Section 8. The storage screen and Rediscover's quick look show their results to you and store nothing.
  • First-look summary: When a new account sees its first-look summary of what the first sort found, SiftMail records that it was shown, which number you tapped and a tap on “Done”, so the summary does not come back. Its counts are ones SiftMail already keeps; showing it reads nothing new from your mail and stores none of it.
  • OAuth tokens: Access and refresh tokens are encrypted by SiftMail with AES-256-GCM before they are stored, using a key kept outside the database (Section 5; a token stored before that encryption was added can remain unencrypted until it is re-saved), and used only to access your mailbox on your behalf.
  • Usage data: Actions taken by you or by SiftMail (quarantine, release, archive, unsubscribe, feedback) with what is needed to undo them, which screens of the dashboard you open, settings, and plan information. Screen names only — never the mail shown on them. Alongside those action records SiftMail keeps the subject of each message a cleanup archived or quarantined, the sender and subject of the most recent messages a cleanup acted on or, in a preview, would have acted on, including the preview the dashboard prepares on its own, and the words you type into SiftMail's message search. When SiftMail shows a free account its Pro offer card, it records when the card was shown and what prompted it (a finished cleanup, a flagged threat or a full Money page), and records a tap on “Not now”, only so the card does not come back too often. The numbers on the card are counts SiftMail already keeps; showing it reads nothing new from your mail. When you share a Rediscover card, SiftMail records that you started a share and how (the share sheet or a download); for the “my year in email” card it also records the card's shape and how many numbers you chose, never the numbers themselves. That card is drawn on your device from totals only, contains no names or addresses, and is not uploaded or kept by SiftMail.

2. How We Use Your Information

  • Score incoming emails for phishing, malware, spam and promotional content, and surface the results in your dashboard. On the Business plan, also check the past week's message records every six hours for senders whose reply-to address changes or whose links point at a lookalike of their own domain.
  • Judge whether each message needs a reply, is safe to archive, or was written by a person, for your dashboard, your digest and recurring maintenance.
  • Build and maintain your Sift Inbox Index — a per-user, searchable, structured index of your mail (receipts, travel, subscriptions, financial, shipping, events, photos, documents, accounts, dining) so you can find buried information using natural language search — and summarise it as Sift Wrapped, a card of counts (how many items, from which years and in which categories) that you can choose to share.
  • Publish counts computed across all SiftMail accounts together on SiftMail's public threat report: how many messages were checked, how many were labelled safe, promotional, spam or phishing, how many needed the AI classifier, how often each kind of warning sign was found, how many passed or failed sender authentication, how many messages people re-marked as safe or promotional, and how many accounts there are. The report contains no message content, sender, subject or address, and names no account.
  • Look for sensitive material sitting in your mailbox (Section 3), and find large or old messages with attachments on the storage screen and in Rediscover (Section 8).
  • Apply a single security label, Sift/Quarantine, in your Gmail when a message is flagged and you or your settings call for it. This is the only label SiftMail creates or applies; the only other label it changes is Gmail's own Inbox label, which archiving takes off and release or undo puts back (Section 4). Your Sift Inbox Index collections (receipts, travel, and the rest) live inside the SiftMail dashboard — they are not Gmail labels and nothing is written to your mailbox for them. Archiving happens in a cleanup you start or, only if it has been switched on for your account, in recurring maintenance on a schedule; SiftMail never archives a message as part of classifying it when it arrives. Section 4 says exactly when each happens and how to undo it.
  • Send you the emails you turn on, and Slack alerts if you connect Slack (Section 6).
  • Authenticate your account and operate features you have enabled in your settings.
  • Pause sifting for free accounts nobody uses. SiftMail notes when you last used it while signed in (your account address and the time, at most about twice a day, kept for 60 days). If a free account goes 14 days without a signed-in visit, SiftMail stops checking, scoring and sending its new mail to the AI classifier, and stops its other scheduled mailbox reads, until the next signed-in visit. Sifting then resumes on its own and checks the mail from that time that is still in your inbox, newest first, up to 500 messages; older mail from that time is not checked. When it resumes, SiftMail also records that it did and roughly how many days you were away, to show you a short notice about it; that record is kept for 60 days. Pausing deletes nothing, releases no quarantined mail and changes none of your settings; routine data clean-up continues on its normal schedule. Pro and Business accounts and trials are never paused.
  • Count, for SiftMail's own operators, how many new accounts reach each step from signing up to paying: connecting a mailbox, the first sort, tapping the first-look summary, seeing or clicking an upgrade prompt, starting checkout and being on a paid plan. These counts come from your account record, your plan and short first-party usage events SiftMail already records (an event name, the time and a few settings such as which prompt it was), never from email content, and they are shown only as totals and percentages. SiftMail uses no third-party analytics for this and sends these events to no one.

3. Exposure Index (Sensitivity Scan)

SiftMail runs an automated sensitivity scan over every incoming message and every message the Sift Inbox Index reads — not only the optional deep scan in Section 8, and not only on paid plans. Its purpose is to show you sensitive information sitting forgotten in your own mailbox. This section describes it in full because it is the most personal thing SiftMail does.

What it looks for. Government identifiers and identity data; credentials such as passwords, API keys, verification and backup codes, and password-reset links; card, bank account and crypto wallet details; financial documents such as tax filings, bank, brokerage, lending and insurance documents, salary and equity details; medical and health records; legal, housing, benefits and employment-related documents and correspondence; children's personal, school and medical records; and confidential work material such as forwarded internal documents, NDAs, source code and third-party personal data you are holding on someone else's behalf.

What it does not do. The scan detects sensitive material that is actually present in a message. It does not profile you, and it does not attempt to infer characteristics you have not written down — SiftMail does not try to determine your health status, pregnancy, sexual orientation, or whether you are looking for a job.

AI inferences. An earlier version of the scan did ask the AI classifier to infer personal circumstances like these from the text of your messages. That instruction has been removed. Inferences it made before then that are still on file stay visible to you in the Exposure Index, so you can see and correct what SiftMail holds about you. Some current AI findings also name no specific kind of sensitive data. Both are labelled "AI inference, not a direct match", because they are the AI's guess from the wording of a message and may be wrong. They are not counted in your exposure numbers and are never included in your digest email or other notifications. You can dismiss any of them in the app, or ask us to delete them (Section 9).

What is stored. For each finding: the type and category, a severity, the sender, subject and date of the message, any attachment filename involved, and a short excerpt of the text surrounding the finding. When the finding comes from SiftMail's own pattern detector, the sensitive value itself is redacted before storage (for example "…SSN ***-**-1234 on file…"). When it comes from the AI classifier, the classifier is instructed to mask the value in the same way, but SiftMail does not check that it did, so an excerpt from an AI finding can contain more of the original text. Where the sensitive data belongs to someone other than you, the finding may record that person's name or role so you can see whose information you are holding.

Who can see it, and how long it lives. Detection is fully automated; findings are stored per-user and are visible only to you, subject to the same human-access restrictions as all other email content (Section 7). Part of the scan uses automated pattern matching only; the remainder uses the AI classification described in Section 6. Exposure findings and the contact directory are deleted along with your other data when you disconnect and do not reconnect, when you delete your account, or immediately on request (Section 9).

4. User Control Over Actions

Quarantine. SiftMail applies one security label — Sift/Quarantine — to messages in your Gmail. You can apply it yourself from the dashboard while Shadow Mode is off. On a paid plan with automatic quarantine turned on and Shadow Mode off, SiftMail also applies it on its own: to new mail it scores as a confident threat, during a threat cleanup you start, and during recurring maintenance to confident threats and to mail from senders you have marked as blocked. That is the only label SiftMail creates or applies (archiving and release, below, change only Gmail's own Inbox label), and it is non-destructive in the strictest sense: quarantining adds a label and nothing else. The message is not moved, not archived, not deleted, and stays in your inbox exactly where it was. You can remove the label at any time from Gmail or from your SiftMail dashboard.

Release. Releasing a quarantined message takes the Sift/Quarantine label off and also adds the inbox label back, so a message you archived in the meantime returns to your inbox. You can release a message from the dashboard. SiftMail also releases one on its own, unless Shadow Mode is on, when its periodic re-check, made by its own rules without AI, scores the message as safe, promotional or from a VIP. That includes a message you quarantined yourself.

Archiving. SiftMail never archives mail on its own except through recurring maintenance, described below, which is off by default. Otherwise archiving happens in a cleanup you start, at any time from the dashboard, which archives inbox mail Gmail files under Promotions, bulk mail that carries an unsubscribe header, and no-reply mail older than 90 days. Before anything is written you are shown a count and have to confirm it — "Archive 13,400 messages?", with that figure counted from your own mailbox. The run then archives every matching message still in your inbox, which can be more than the figure you confirmed. It runs in bulk, not one message at a time, which is precisely why the number is put in front of you first. Archived mail stays in Gmail under All Mail and search still finds it, and the entire run can be reversed from the Cleanup tab. Archiving frees no Google storage.

Recurring maintenance is a separate, opt-in setting, off by default, and the dashboard does not yet offer a control to turn it on. It acts only for a paid account, and only once it has been switched on and taken out of both its own preview mode and Shadow Mode; in preview it reports what it would do and changes nothing. Once it acts, it runs on a schedule without asking again each time — that is what turning it on means. Each run archives recent inbox mail that SiftMail's classifier is confident is safe to archive: promotional, bulk and no-reply mail (no-reply here includes notification, update and newsletter senders), automatic replies such as out-of-office notices, and any other mail the classifier rates as safe to archive. It never archives starred or important mail, mail that appears to need a reply, mail the classifier judges a person wrote, mail from a sender in your contact directory or one you have marked as trusted, or mail from your own domain. Each run records an undo entry, and SiftMail emails you a summary of these runs, as Section 6 describes. A summary carries an Undo link for one run, the latest, when that run changed mail; a weekly summary's link does not cover the week's other runs. The link opens the dashboard, which shows that run's counts and, only once you confirm there, puts the inbox label back on what the run archived and takes the quarantine label off what it quarantined. That link is the only way to undo a maintenance run: a run no summary links, because the summary is turned off, because a later run is the one a summary linked, or because the mailbox was no longer connected when its summary was due, has no undo control yet. A later run does not archive again a message an undo put back. It does not quarantine a message that already carries the quarantine label, or one taken out of quarantine, whether by an undo, by you or by SiftMail's own re-check; a message taken out of quarantine can still be archived later, like any other mail a run archives.

Trash. One screen can move mail to Trash: the storage screen, whose whole purpose is to free Google storage, which archiving does not do. It is never automatic, never triggered by a classification, and never part of a cleanup or a maintenance run — you pick the messages yourself and confirm each batch, capped at 200 at a time. At the moment you confirm, SiftMail re-reads every one of those messages from Gmail and refuses any draft or anything you have starred, even if it looked selectable when the page loaded. Mail moved to Trash stays in Gmail's Trash for 30 days, where "Move to inbox" puts it back; that is why this is the one action SiftMail does not record its own undo for.

Unsubscribe. If you ask SiftMail to unsubscribe you from a mailing list, SiftMail's server opens the unsubscribe web address that the sender put in that message's headers, which tells the sender to stop, and records the address and whether it worked. SiftMail never sends an unsubscribe email for you; for a list that only accepts one, it gives you the address to write to yourself.

Your allow, block, VIP and trusted-sender lists change only through something you do: editing them, or marking a message safe, VIP, spam or phishing, which can add its sender to the matching list. You can disconnect SiftMail at any time from your dashboard or from your Google Account permissions page.

5. Data Storage & Security

  • SiftMail's stored data is kept in a PostgreSQL database. Work waiting to run passes through a job queue, which also keeps a limited number of recently finished and failed jobs for each kind of work; those can carry results such as the senders a digest counted, or an error message. Both are hosted on Railway (Section 6), and both need a password to connect to.
  • SiftMail itself encrypts a few items before storing them in the database: the OAuth access and refresh tokens that let it reach your mailbox, the value of each saved fact (see Section 1) and, if a Slack connection is ever set up, its webhook address. All of these use AES-256-GCM. The keys for this are kept in SiftMail's server configuration on Railway, separate from the database, so someone holding only a copy of the database cannot read a token or webhook address saved since that encryption was added. One saved before it, on an older account, can still be in the database unencrypted until it is replaced: a Slack webhook address in an older settings record, or an OAuth token that has not yet been re-saved. Saved facts are protected less completely, in two ways. Each one's last four characters are stored unencrypted for the masked list, so a fact that is only four characters long, such as a card or account ending, can be read in full, and a five-character one all but its first character. And a saved fact is found in a message's subject or Gmail snippet, which SiftMail keeps unencrypted in its message records, so the same value can usually also be read there for as long as that message record is kept.
  • Everything else SiftMail stores, such as message records, verdicts, settings and the action history, does not get that extra layer of encryption from SiftMail. It is protected by those passwords and by Railway's own access controls.
  • Traffic between your browser or the Chrome extension and SiftMail, and SiftMail's own calls to Google and to its AI provider, are encrypted in transit with TLS (HTTPS).
  • SiftMail never writes a full email address to its application logs: wherever one would appear, only the first four characters before the @ are kept and the domain is dropped. Sign-in codes and access tokens are removed, web addresses are logged without the values in their query string (which is where search terms travel), and the IP address recorded with each request is cut to its network prefix. Railway's edge request logs may record the path of a calendar-feed request.
  • We use rate limiting and API key authentication to protect endpoints.

6. Third-Party Services

  • Google OAuth and Gmail API — to authenticate you and access your Gmail mailbox.
  • Microsoft Identity Platform and Microsoft Graph API — to authenticate Outlook users and access their mailbox.
  • Anthropic API — to perform AI classification of email content, in two separate checks; one message can go through both. Security scoring is used only for messages our own scoring cannot resolve confidently. It sends the sender, the subject, up to 600 characters of Gmail's short preview snippet, the message's sender-authentication results, and the risk signals our own rules found, which can quote values from its headers such as a reply-to domain or a display name. When the first model is unsure or its answer is borderline, the same content is sent again to a second Anthropic model. Inbox indexing, the sensitivity scan in Section 3, and the reply and archive signals are one check. It sends the sender, subject, date, attachment file names and types, and up to the first 4,000 characters of the message body, for new inbox mail and, once you start the Sift Inbox Index, for past mail; mail our own rules already recognise as marketing, or can file by themselves, is not sent. Mail SiftMail recognises as legal or health-related is not sent to Anthropic for indexing either: it gets no AI categories, contacts or AI-found exposures, while SiftMail's own pattern detector still checks it for things like identity numbers and card details (Section 3). That recognition is automated pattern matching, so it can miss some legal or health mail. Security scoring is not covered by this exclusion: a legal or health message our own scoring cannot resolve can still have its sender, subject and short preview snippet sent to security scoring as described above, because phishing often poses as court or medical notices. A message is not limited to a single check of each kind: it can be sent again, for example when a later sync picks it up again, when an earlier attempt failed, or when the index reaches a message an earlier check filed nothing for. When you search your index in your own words, only the words you typed are sent. The Rediscover deep scan in Section 8 sends no message content to Anthropic at all. Nor do Money, Life admin, Ask, reply drafts or follow-ups (Section 1). Email content sent to Anthropic is processed under Anthropic's API terms, is used solely to return classification results, and is not used to train AI models.
  • Stripe — for subscription billing and payment processing. Stripe receives your account email address and the plan you buy, and no mail content. SiftMail does not receive or store your full payment card details.
  • Resend — for the emails SiftMail sends you. They go to your account's address, except the maintenance summary, which goes to the mailbox named in your maintenance settings. That can only be your account's address or a mailbox you have connected to your account, and SiftMail checks it again before each summary and sends none to a mailbox that is no longer connected. These are: the daily or weekly digest, on a paid plan if you turn it on (once the free digest is switched on, which it is not yet, a free account that turns it on gets a lighter version with the same counts and up to three messages), listing the senders and subject lines, risk scores and arrival times of the messages it summarises, with a link to each in Gmail, and, for flagged messages, the From and Reply-To addresses that triggered the flag; once the Monday brief is switched on, which it is not yet, a Monday digest for a Pro or Business account also includes a short “Your week ahead” block derived at send time from the same message records SiftMail already stores, based on about the last 90 days (not a complete mailbox history) — threads waiting on a reply (who and subject, with a link to open the source conversation in your mailbox when SiftMail has a usable mail id), suggested things you may have said you would do, drawn only from Sent messages that have already been imported, as a short matched clause of up to 120 characters from Gmail's snippet of that sent message (not the full body; these are pattern-matched suggestions, not confirmed to-dos), how many open sensitive-content findings you have among the stored findings checked (up to 10,000) and whether that count changed this week, and how many new correspondents appeared this week among contacts seen in that 90-day window — and never full message bodies; a summary after recurring-maintenance runs, at most about once a day, sent by default once recurring maintenance is switched on, including while it is in preview, which gives counts, up to five messages that need a reply with their senders, subject lines and up to 120 characters of Gmail's snippet, and up to three quarantined messages with their senders, subject lines and the reason, and, for flagged messages, the From and Reply-To addresses that triggered the flag; a weekly note, on a paid plan with the digest turned on and only in a week when your sensitive-content findings changed, giving the kinds of finding, their severity, the year and counts, and never senders, subjects or excerpts; and, if you subscribe to it, a promotional-mail digest listing the senders, subject lines, categories and dates of promotional messages.
  • Slack — only if you choose to connect it. When you supply your own Slack incoming-webhook URL, SiftMail can post two things to it: a connection test when you ask for one, which carries no mail content, and an alert when a message checked through SiftMail's single-message check gets one of the labels you choose — phishing and spam by default, and promotional mail too if you tick that box. The dashboard and SiftMail's own checks of your new mail do not use that single-message check, so they do not post to Slack. Each alert contains the sender, subject line, risk score and label, and the signals behind the score, which can include the AI classifier's short written assessment of the message. No data is sent to Slack unless you configure this yourself, and you can remove the webhook at any time.
  • Have I Been Pwned (Pwned Passwords) — for Pro and Business accounts, when SiftMail finds a password in your mail. It receives up to two k-anonymity range queries per detected password, at most once a day: the first 5 hex characters of the SHA-1 of the password as written, padded, and, when that value ends in sentence punctuation, a second such prefix for the punctuation-trimmed spelling — and nothing else: no password, no full hash, no email address, no message.
  • The owner's AI assistant service — only when SiftMail's owner runs an account search (Section 7) through it. The provider of that assistant receives the account details the search returns: email address, plan and billing status, sign-up date, connected mailboxes with their status and last sync time, sign-ins, when the account was last used, and its SiftMail settings. No email content, subjects, senders or message counts reach it this way.
  • Railway — for application hosting, the database where everything SiftMail stores is kept, and the job queue that carries work between SiftMail's servers, such as which message to check next and for which account.

Each provider is governed by its own privacy policy. Apart from these services, data from your mailbox goes to two other places. If you ask SiftMail to unsubscribe you from a mailing list, its server opens the unsubscribe web address the sender put in that message (Section 4); that happens only when you ask. And SiftMail's public threat report (Section 2) shows counts added up across all accounts together, never a message, a sender or an account.

7. Google API Limited Use Disclosure

SiftMail's use and transfer of information received from Google APIs adheres to the Google API Services User Data Policy, including the Limited Use requirements.

  • We request the gmail.modify scope because SiftMail's core functions require both reading message content (for security classification and inbox indexing) and writing labels. Quarantine adds the Sift/Quarantine label and leaves the message where it is, and releasing a message takes that label off and adds the inbox label back. Archiving is a separate action that happens only when you start a cleanup, or on a schedule once recurring maintenance has been switched on for your account: it removes the INBOX label and nothing else; a cleanup runs in bulk once you confirm the count shown, and recurring maintenance, which is off by default, archives without asking each time — section 4 describes both — and each writes an undo entry. The same scope powers Trash, which you explicitly confirm, which is capped at 200 messages per call, and which sends mail to Gmail's own Trash. A read-only scope cannot write labels; a metadata-only scope cannot power the Sift Inbox Index.
  • We do not use Google user data for serving advertisements.
  • We do not allow humans at SiftMail to read your email content except in narrow circumstances expressly permitted by the Google API Services User Data Policy: (a) with your explicit written consent to investigate a specific issue you have reported, (b) to comply with applicable law, or (c) to investigate suspected abuse or security incidents.
  • SiftMail's owner can search accounts by the start of an email address (at least five characters, or at least three followed by @) to run the service and answer support questions. A search returns at most five accounts and shows only account details: the account's email address, plan and billing status, sign-up date, connected mailboxes with their connection status and last sync time, sign-ins (how many and when, only as far back as SiftMail's audit records go: 180 days for free accounts, 365 for paid), when you last used SiftMail, and your SiftMail settings. It never shows email content, subjects, senders, counts of your messages, or anything about the people who write to you, and it changes nothing in your account. Each search is recorded in the audit records of the owner's own SiftMail account, with the text searched for, the number of accounts found and the IP address the search came from; that record is kept for 180 days, or 365 while the owner's account is on a paid plan. The owner may run these searches through an AI assistant service, so the account details a search returns are processed by that assistant's provider (Section 6).
  • We do not transfer Google user data to third parties except as necessary to provide the service — to the services named in Section 6, and to a mailing list's own unsubscribe address when you ask to unsubscribe (Section 4) — or as required by law. Separately, SiftMail publishes counts added up across all accounts on its public threat report (Section 2); they contain no message content and name no account.
  • All Google user data access is limited to the practices explicitly disclosed in this privacy policy.

8. Rediscover Deep Scan (Decade Audit)

SiftMail offers an optional deep scan — surfaced as Rediscover and sold as the one-time Decade Audit — that goes through the messages in your mailbox that have attachments, year by year as far back as 2004, when Gmail launched, to surface old media, receipts, and forgotten sensitive documents you may not remember are still in your mailbox. Rediscover also offers a quick look, which reads a page of messages with attachments, of any age, when you ask, shows them to you and stores nothing. When you open Rediscover it first shows a summary of your past year (recurring charges, refunds still owed, return windows still open on recent orders, announced price increases, warranty and protection-plan emails, accounts and people you have gone quiet with, and mail from about a year ago). It is built only from the message records SiftMail already keeps (Section 2), by SiftMail's own rules without AI; it reads nothing new from Gmail, stores nothing, and sends nothing to anyone.

  • What we read: For each message, the deep scan fetches the full message from Gmail, which includes its body, to find its attachments. It uses only the message's headers, the names, types and sizes of its attachments, and Gmail's short preview snippet; it does not use or store the message body, and it never downloads attachment files.
  • What we store (locate, not custody): For each matched message the deep scan persists metadata only — sender, subject, date, category, attachment metadata (filename, type, size and count), whether it was classified sensitive and, if so, the kind and severity of what was detected with a short masked note, whether SiftMail's own rules flag it as likely old phishing or spam and why, and a message locator (the Gmail message ID) so you can jump back to the original. SiftMail does not copy or store the full message body or the contents of your attachments from the deep scan.
  • Sensitivity classification: The deep scan flags forgotten sensitive documents and records the sensitivity class it detected — including financial (e.g., tax returns, bank/brokerage statements, pay stubs), identity (e.g., passports, driver's licenses, government IDs, Social Security numbers), and medical (e.g., lab results, prescriptions, insurance cards), as well as legal and location-related documents. Detection is fully automated (filename, subject, and metadata patterns, and for the phishing check Gmail's snippet); no human at SiftMail reviews your documents.
  • Access & retention: Deep-scan findings are stored per-user, are visible only to you, and are subject to the same Limited Use and human-access restrictions described in Section 7. They are deleted when you disconnect and do not reconnect (within 30 days, per Section 9), when you delete your account, or immediately on request to privacy@siftmail.app.

9. Data Retention & Deletion

You can disconnect your account at any time, which immediately revokes our access to your mailbox and stops further processing of your email. When you disconnect and do not reconnect, your cached email data, Sift Inbox Index, Rediscover deep-scan findings, and scoring history (including the scoring records in Section 1) are deleted within 30 days. Confirmed contact merges, reply drafts, follow-ups and snoozes (and vouched senders, Not a reply senders and saved facts) are deleted sooner: within 24 hours of your last connected mailbox being disconnected. Entries in your action history are kept as a record that each action, including the deletion itself, took place, but the message details in them are cleared at the same time as the rest of your data. Separately from disconnecting, reply drafts are deleted 30 days after they are created, and finished follow-ups 90 days after their due time. When you delete your account, all of this, including the action history, is deleted. You may also request immediate, complete data deletion at any time by emailing privacy@siftmail.app.

10. Your Rights

You have the right to access, correct, or delete your personal data. You can disconnect your email provider at any time through the dashboard. On the Trust screen you can download a copy of your confirmed contact merges, vouched senders and Not a reply senders, and clear them, yourself. For any other data export or deletion request, contact us at privacy@siftmail.app.

11. Contact

For privacy-related questions or requests, email us at privacy@siftmail.app.