The latest tagged release only. Fixes land on main and ship in the next release; older tags are not patched in place.

Use GitHub's private vulnerability reporting: the repository's Security tab → Report a vulnerability. That opens a private advisory visible only to you and the maintainer, so a working exploit can be discussed before there is a fix to point people at.

No email address is published for this, on purpose — a private advisory is tracked, attributable and has a disclosure timeline attached; an inbox has none of those.

Please include what you would put in a bug report: version, front-end, and the exact steps. A proof of concept is welcome and never required.

By default, AutoShade keeps the source library read-only. If the configured Delivery folder is inside or above a photo’s folder, that delivery subtree is intentionally writable; Settings warns when this removes the folder’s protection. “Export .xmp beside the photo” is the separate, confirmed per-photo sidecar exception.

The parts with a real attacker are the ones that take input from somewhere else: the local web server (serve), the settings and key-resolution rules, the child processes (the claude verifier and the Python sidecars), and the third-party decoders that parse untrusted RAW files. Both threat models are already written down — the server's in User manual § Privacy, trust, and paid-feature boundary and ARCHITECTURE § 4.9 "What the local web server refuses", the settings-trust boundary in the same manual section. If you have found a way past what those documents claim, that is exactly the report worth making — and so is a case where the documentation overstates the guarantee.

Out of scope: anything that requires an attacker to already be running code as you, and the behaviour of the AI endpoints you configure — those are third-party services under your own account.