MLXRead is a local macOS app that reads selected text aloud with an on-device model. Its core security promise is narrow and absolute: selected and spoken text must never leave the Mac. The public website, support form, delivery Worker, signed app, and update path are also part of the security boundary.
Please report privately first; do not open a public issue for a security bug.
- In-app: Settings → Report sends a diagnostic bundle and optional reporter-entered email/description to the maintainer. The bundle must never contain selected or spoken text.
- Web: use mlxread.com/support and choose
“Privacy.” The temporary
mlxread-web.pages.devhostname may remain available for deployment compatibility, butmlxread.comis the canonical surface. - For sensitive reports, request an encrypted channel through either route.
Include the affected version, macOS version, and reproducible steps. Do not paste selected or spoken text into either report channel. Expect an acknowledgement within a few days.
In scope:
- The macOS app: selection capture, the Option–Escape event tap, clipboard fallback, audio, model loading, local storage, and privilege boundaries.
- Release and update integrity: Developer ID signing, notarization, Sparkle appcast configuration, and EdDSA verification.
- The unauthenticated public website (
website/) and delivery Worker (api/), including contact/report submission, report storage, and download links. - Build and dependency paths that could compromise the signed artifact.
MLXRead has no accounts, tenancy, analytics, or application cookies. Website support messages and in-app reports are nevertheless user data: permitted fields include reporter email, message/description, app/build/macOS/hardware metadata, settings state, request IP metadata, and privacy-safe app logs.
Out of scope:
- A third-party dependency's own advisory when there is no demonstrated MLXRead-specific reachability or impact; report those upstream.
- Attacks that assume an adversary already controls root or an administrator account and cross no additional MLXRead boundary.
Same-user unprivileged attacks, local privilege escalation, signing/update bypass, report-link disclosure, and violations of the selected-text boundary remain in scope.
- Reading content stays local. Selected text is synthesized in-process, held only while needed, never written to disk, never placed in website or report payloads, and never logged. See docs/privacy.md.
- Reports are deliberately bounded. Debug bundles contain enumerated metadata and recent MLXRead logs only. The Worker enforces request and bundle size limits, best-effort throttling, and a mail binding restricted to the maintainer. User-entered descriptions are not automatically populated from a selection and must not include reading content.
- Public report storage uses bearer capabilities. R2 objects are addressed
by random 128-bit identifiers and returned with
private, no-store; anyone who obtains a download URL can read that bundle. Retention and authenticated maintainer access are tracked limitations, not protections that exist today. - Release artifacts are signed. The manual-install DMG and its contained app must be Developer ID-signed, hardened, notarized, and stapled. The ZIP is retained as the separately verified Sparkle update archive. Debug builds deliberately have a weaker debugger posture and are not release artifacts.
- Updates fail closed when unconfigured. Sparkle starts only with a real HTTPS feed and non-placeholder public key, then verifies EdDSA signatures. The checked-in feed and project settings remain templates, so source builds do not currently establish an active update channel.
- The public web surface is stateless. It uses no accounts or cookies and applies a CSP, frame denial, content-type restrictions, same-origin checks, and body-size limits to server-function requests.
The repository-specific assessment is maintained in mlxread-threat-model.md.
Confirmed app vulnerabilities are fixed in a new signed and notarized release. When an update channel is configured for that release, its signed appcast is updated as part of publication. Website and Worker fixes require their own deployment and live readback; a source change or local build is not proof that production changed. Credit is given in release notes unless the reporter asks to remain anonymous.