birdnet is under active development and ships a single line of releases. Security fixes are made against the latest release on PyPI. Please upgrade to the latest version before reporting a problem. We do not backport fixes to older releases.

Please do not open a public issue for security problems.

Report privately through GitHub's Report a vulnerability button (Security tab → Report a vulnerability). If you'd rather use email, contact josefhaupt@gmx.net.

Please include enough to reproduce: the affected version, backend/model, platform, and a minimal proof of concept. You can expect an initial response within a few days — this is a small, maintainer-run project, so please allow reasonable time for a fix before any public disclosure.

The parts of this library most relevant to security are less about the inference math and more about what gets downloaded and loaded from disk:

• Model and taxonomy auto-download. Official models and the V3.0 taxonomy are fetched over the network (~3 GB) on first use. Issues in how those artifacts are retrieved, verified, or cached are in scope.
• Loading model files. Loading a model can deserialize third-party files — notably torch.load (pickle) for .pt backends, and SavedModel/TFLite artifacts. Loading a custom or untrusted model file (load_custom) executes whatever that file contains; treat model files like code and only load ones you trust. Vulnerabilities in our loading path are in scope.
• Path and file handling for audio inputs and result exports.

• crashes on malformed audio (please file those as a regular bug), model prediction accuracy, and vulnerabilities in optional dependencies themselves (report those upstream, though we're glad to bump a pin).