Report privately through GitHub Security Advisories — the Security tab of this repository, then Report a vulnerability. Please do not open a public issue for a security problem.

You can expect an initial reply within a few days. This is a small project maintained in spare time, so please be patient with follow-ups.

The tool shells out to ffmpeg, ffprobe, whisper-cli and yt-dlp, and passes user input (file paths, URLs, option values) to them. It also invokes agent CLIs non-interactively with permission to write files in the working directory. Problems worth reporting include:

• command or argument injection through a file path, URL or option value
• a path traversal that writes outside the intended vault directory
• an agent invocation that grants broader permissions than documented (acceptEdits on Claude, workspace-write on Codex, auto_edit on Gemini)

--cookies-from-browser and --cookies hand yt-dlp a real session. A YouTube cookie is your Google account: treat it as a password, not as configuration.

• The account carries the risk, not just the file. Cookies attach automated requests to your account rather than to an anonymous IP, and downloading is against YouTube's terms regardless — signed in, it is attributable to you. Suspensions after heavy use are reported often enough in the yt-dlp community to take seriously. Use a secondary account, not your main one.
• Prefer --cookies-from-browser. It reads the browser's store for that one request and writes nothing. On macOS the browser will ask for Keychain access — that prompt is expected.
• A cookies.txt is a live credential on disk. It is git-ignored here, but it is still a file anyone with read access to that directory can use. Delete it when you are done, and never commit or share one.
• Mind the agent. --claude, --codex and --gemini run an agent with write permission in the working directory, so anything readable there is readable by it. That is one more reason not to leave a cookies.txt beside your vaults.
• The cookies are passed to yt-dlp as arguments and never written to the vault, the logs or meta.json.

• Vulnerabilities in ffmpeg, whisper.cpp or yt-dlp themselves — report those upstream.
• The fact that the tool does not work around DRM or authentication. That is deliberate, and reports asking for it will be closed.
• Content processed by the tool staying on your machine is the design; there is no server and no telemetry. The network is touched in exactly three places: yt-dlp fetching a public video, the one-time Whisper model download, and the rendered HTML page loading its typefaces from Google Fonts — that last one at reading time, in the reader's browser, never while a video is processed.