SoftVC VITS Singing Voice Conversion contains a path traversal vulnerability in the full‑song inference server that allows unauthenticated attackers to read arbitrary files by supplying an attacker‑controlled audio_path in a POST request to the /wav2wav route. The server passes the supplied path directly to librosa.load, torchaudio.load, and soundfile.write, enabling the attacker to retrieve any file contents in the HTTP response and also to write arbitrary .wav files to any location on the filesystem. This flaw provides remote file disclosure and arbitrary file write capabilities.

The vulnerability exists in any release that includes commit 730930d of the SoftVC VITS project maintained by svc‑develop‑team under the so‑vits‑svc product line. All versions of the full‑song inference server that have not applied the mitigation for this commit remain affected, regardless of the overall SoftVC VITS package version.

With a CVSS score of 9.3 the flaw is classified as critical, and its EPSS score of less than 1 % indicates a low but non‑zero probability of exploitation at the time of analysis. The attack surface is network‑based and unauthenticated; any external user can submit a crafted POST request to the vulnerable endpoint. Because the attacker can read sensitive files such as configuration, credential, or personal data and can also write files that may overwrite or inject malicious code, compromise of the server could lead to data leakage, credential theft, or serve as a foothold for deeper attacks. The flaw is not currently listed in CISA’s Known Exploited Vulnerabilities catalog.