Only the latest release gets fixes. Check yours under About, or against the releases page.

Please use GitHub's private reporting — Report a vulnerability — rather than a public issue, and give it a few days before disclosing.

Helpful to include: the version, the OS, what an attacker would gain, and the smallest file, URL or configuration that reproduces it.

Windows builds from v1.3.9 through v1.5.0 shipped a maintainer-owned Google Cloud service-account key at creds/gcloud_stt.json inside the package, so that Google Cloud Speech-to-Text worked without any setup. Because a key was present, it also became the default engine — a fresh install transcribed through the maintainer's cloud account unless the user changed it.

Why that was wrong: a credential inside a public download is a credential handed to everyone who downloads it. It could be extracted from the package and used outside the app entirely; all usage was billed to one project; and that project's owner was answerable, under Google's terms, for whatever anyone chose to transcribe with it.

What was done:

• The key was revoked by the maintainer. Any copy extracted from an old build is dead.
• Both local copies were deleted, and the build steps that copied a key into the package are removed — build_embed_installer.bat now errors if a key is sitting there, and neither PyInstaller spec adds one to datas.
• The default engine is unconditionally offline faster-whisper. Neither core.config._default_transcribe_backend nor core.backends.availability.default_engine looks at a bundled key any more, so dropping a key next to the app cannot change anyone's default.
• Regression tests in tests/core/test_engine_selector.py fail if either behaviour comes back.

If you run v1.5.0 or older, cloud STT will stop working; pick your own service-account JSON in Advanced > Backend (docs/CLOUD_STT_GOOGLE.md). Nothing else in the app is affected — offline transcription never used the key.

Do not re-introduce a shared credential. If a hosted service is ever wanted, put it behind a server the project controls, with per-user quota, rather than a key inside the download.

This is a desktop app that reads local media files, shells out to ffmpeg, ffprobe and yt-dlp, runs the speech model in a subprocess worker that talks newline-delimited JSON over stdin/stdout, and — only when the operator turns it on — serves an HTTP endpoint on the local network. Everything runs with the rights of the user who launched it.

• A crafted media file, transcript file or URL that escapes into command execution, or that makes the app write outside the chosen output folder and its own data directory.
• Anything that lets a Web / LAN access client read or write files outside the intended upload/output paths, bypass the access password, or reach the host beyond the documented routes.
• Leakage of the credentials used by the two opt-in cloud backends (a pasted Gemini API key, a Google service-account JSON) into logs, crash reports, transcripts or the update check.
• Tampering with the update check so it points a user at something other than this project's own releases.
• Privilege or persistence surprises from the installer.

• Vulnerabilities in ffmpeg, yt-dlp, faster-whisper/CTranslate2, the bundled Python or the model weights themselves. Report those upstream; see THIRD_PARTY_NOTICES.md.
• Web / LAN access used on an untrusted network. It is documented as trusted-network-only: no accounts, no TLS, an optional shared password. Anyone who can reach the address can use it, by design. Exposing it to the open internet is a deployment choice, not a bug.
• The two cloud backends uploading audio. That is their whole purpose, they are off by default, and both the UI and the README say so before you enable them.
• Antivirus heuristics flagging the installer or the packed executable.