Email support@ohmail.app with SECURITY in the subject. That address is the
one published on ohmail.app and is monitored by the team at
TrafficFlow GmbH; there is no separate security alias yet, and we would rather
say so than publish an address nobody reads.
Please include: what you found, how to reproduce it, the version or commit you tested, and what you think the impact is. If you have a patch, send it in the mail rather than opening a PR.
Do not open a public issue for a vulnerability. Everything else — crashes, layout bugs, wrong copy — belongs in an issue.
We will acknowledge within 5 working days, tell you our assessment and a rough timeline, and credit you in the release notes if you want the credit. We do not run a bug bounty. We will not threaten you for reporting in good faith.
In scope: this repository — the macOS client, the Windows/Linux Tauri shell
(apps/desktop, including its CSP, its capability set and the offline guard),
the shared client UI they render, the build and packaging scripts, and the CI
workflow.
Out of scope here, but still worth reporting to the same address: ohmail.app, app.ohmail.app, and the ohmail Cloud backend. None of that code lives in this repository.
The same address, the same commitments, and the same 5-working-day acknowledgement cover ohmail.app, app.ohmail.app and the Cloud backend — the code is not published, but the disclosure policy is, and reports about it are as welcome as reports about this repository. Two limits, stated so nobody has to guess: please do not test against another person's account or mailbox, and please do not run load or denial-of-service tests against the hosted service. Cloud mail is encrypted in transit and at rest but is not end-to-end encrypted — the service has to read mail to file and search it. The subprocessor list and the retention table are published at https://ohmail.app/subprocessors; the conditions under which a human at TrafficFlow can reach production data are written down in the product privacy policy, which publishes before the first real mailbox connects. If personal data is ever breached we notify the competent authority within 72 hours of becoming aware, and affected people directly where the risk to them is high.
Worth knowing before you go looking, because the two builds have very different threat surfaces right now.
macOS — the engine-bearing build — connects to your mailbox (and, if you ask it to, to ohmail Cloud):
- In local mode it opens one IMAP connection over TLS to the mail server you give it, and — only if you enable it — one connection to your own Anthropic key or a local Ollama. Beyond that it makes the signed update check, fetches the pictures in a message you open (a switch makes each message ask first), sends the one-click unsubscribe a screen-out asks for where the list offers one (a switch turns it off), and, where your provider does not confirm senders itself, looks up through your system's DNS resolver the key a company or a newsletter signs its mail with. There is no telemetry and no analytics.
- In Cloud mode — an optional sign-in — it connects instead to
api.ohmail.app, the hosted ohmail service, over HTTPS, and acts as a viewer of a mailbox that service already organises. It runs no local engine and opens no IMAP connection in this mode. The Cloud session is held in memory only and re-established each launch; no mailbox credentials are stored on the device. - The packaging step reads every host string in the shipped binary and fails the build on
any host outside an explicit allow-list —
api.ohmail.app, the signed update feed, and the local-engine pipe — so a telemetry or third-party endpoint cannot ship unremarked. - Credentials: your mail password is sealed under a per-install AES key and is never
written to disk in the clear. That key is kept in your computer's keystore — the login
Keychain on macOS — and mirrored to a
0600file beside the app's data, because an application without a developer certificate is refused its own keystore item as soon as its binary changes, which is every update. Where the keystore refuses, the file is the only copy: the key is then protected by file permissions rather than by your login password, and it sits beside a local mail mirror that is an ordinary unencrypted database. On Linux the keystore is the Secret Service; on a session that has none, the file is the only copy from the first launch, and a locked keyring stops the start rather than being stepped around. Reports about the credential envelope, the key file, the keystore use, or TLS/certificate handling are exactly what we want. - It renders mail as HTML and blocks remote content, so the tracking pixels that report when and where you opened a message do not load. HTML-rendering and remote-content issues are in scope.
- Sensitive mail — one-time codes and login links — is structurally excluded from anything an AI provider sees: a protected message has no body field to leak, and it is stored redacted.
- The organiser is single-writer per mailbox by design: exactly one copy of ohmail organises a given mailbox at a time, arbitrated through a claim in the mailbox itself, not through any server.
The desktop window itself reaches nothing — networking belongs to the processes around it:
- The webview makes no network connection at all, enforced three times over
and each worth attacking separately: its CSP's
connect-srcnames one source, the app's in-process command scheme in each platform's spelling (apps/desktop/src-tauri/tauri.<os>.conf.json), which on Windows CSP also matches ashttps://ipc.localhost, a loopback address nothing serves; the page replacesfetch,XMLHttpRequest,WebSocket,EventSourceandnavigator.sendBeaconwith functions that throw (proven on the built bundle by the render check's offline audit); and its whole reach is a small named list of shell commands (WINDOW_COMMANDSinapps/desktop/src-tauri/build.rs). None of those commands takes a filesystem path, one takes URLs, and none opens an address the window names.rendered_links_batchlets the window's renderers report the http(s) links they draw (a message's links, a draft's or a signature's, the Screener's unsubscribe link, the account page the service named) and which of a message's links have text naming a different site; the shell keeps the links in a bounded list and opens nothing for them. An address the window reports as marked asks before it opens, wherever you press it and whether or not the list kept its link, until the app quits; the shell keeps up to 1 MiB of such addresses. Past that, a message whose marked address found no room makes every link ask until the window reports that message again or takes it away, and more than 1,024 such messages at once make every link ask until the app quits. When you press a link that asks, the shell holds the press and the window asks you first;link_ask_opentakes only the random token the shell handed out with that question, and the shell then opens the one link it holds, once, within two minutes of your press. A link reaches your browser only as the webview's own new-window request, and only when the webview itself marked that request as your click or key press (apps/desktop/src-tauri/src/link_gate.rs): WebKitGTK'sis_user_gestureon Linux, WebView2'sIsUserInitiatedon Windows, and on macOS WKWebView's new-window question, which it asks only during a gesture because the window is built withjavaScriptCanOpenWindowsAutomaticallyoff. The address must then pass the shell's own vetting (external_urlinengine.rs) and be one of those reported links or one of the shell's own constant links (apps/desktop/src-tauri/src/rendered_links.rs) — every other outbound address the shell can open is a named constant. What remains is narrow, and we state it: a script already running in the page can open a link on screen at the moment you act — on macOS and Linux only inside that click's or key press's own handling (under about a second, and more than once per action), on Windows within Chromium's activation (about five seconds, once per action) — and never without an action of yours. Beside the three, every window the shell builds is locked to the app's own pages: a navigation to any other address is refused and goes nowhere, on every platform (apps/desktop/src-tauri/src/navigation.rs), so a script that got into the page cannot take the window to a page none of the three governs. What connects is the mail engine — to your own mail server, and on the hosted door to your account — and the Rust shell's updater, to its one pinned signed feed. One other window has a grant — the transient one shown while an update downloads — and it is a single receive-only event permission (core:event:allow-listen) so that it can hear the local download-progress event: no emit back, no command, no filesystem, and a page whose own CSP isconnect-src 'none'. - Because the interface is embedded uncompressed, you can audit a downloaded
binary directly:
strings -a <binary> | grep -oE 'https?://[^ ]+' | sort -u.
Both: the CI-built artifacts are unsigned — ad-hoc signature only on macOS, no Authenticode on Windows, nothing on Linux. That is a distribution weakness we name openly in the README rather than a vulnerability to report: verify the artifact came from the CI run you expect, or build from source.
The project is a pre-1.0 preview. Only main is supported; fixes land there.