CVE-2026-65701 Details

A path traversal vulnerability has been identified in the SoftVC VITS Singing Voice Conversion full-song inference server, specifically in versions through commit 730930d. This vulnerability allows unauthenticated remote attackers to read and exfiltrate arbitrary files by sending attacker-controlled filesystem paths through the audio_path field of an unauthenticated POST request to the /wav2wav route. The server decodes the specified paths using librosa.load, torchaudio.load, and soundfile.write, returning the file contents in the HTTP response while also writing .wav files to locations specified by the attacker.

Do not accept server-side filesystem paths from requests. Instead, accept audio files as uploads via request.files, similar to the existing functionality in flask_api.py for other routes. If it is necessary to accept paths, ensure they are canonicalized and confined to a designated input directory, rejecting URLs and non-allowlisted file extensions. Additionally, bind the server to 127.0.0.1 by default, implement an authentication token, and remove the wildcard CORS policy. Finally, write converted files to a controlled temporary directory instead of alongside the source file.