Security fixes target the current main branch and the latest tagged minor release. Older release lines receive fixes only when explicitly announced.

Please report security issues privately to Sonic Field Labs before public disclosure. Include the affected commit, local configuration, reproduction steps, and whether generated audio, model files, or listening records can be exposed.

GERM is a local sidecar. Its optional Python provider loads model and LoRA files only through configured model roots. Treat every model artifact as executable input: use the official Safetensors releases, verify provenance, and do not load untrusted pickle-based checkpoints.

The default server binds to 127.0.0.1, validates Host headers, and rejects foreign browser origins for state-changing requests. Audio and metadata routes normalize paths and require resolved files to remain inside their configured input, output, metadata, model, upload, or scratch roots before filesystem access. Upload writes are confined to managed upload or scratch directories.

Stable Audio 3 still pins PyTorch 2.7.1 upstream. GERM overrides that constraint with Torch and Torchaudio 2.10.0, validated on macOS by the full GERM suite and the upstream Stable Audio CLI suite. This removes every currently fixable finding below PyTorch 2.10.

Two findings remain accepted temporarily for the optional local provider:

The exception ends immediately if GERM begins calling either API, if its model trust boundary changes, or when a compatible upstream runtime is available. The all-extras CI audit ignores only these identifiers and fails on any new finding.