CVE-2026-65701 in so-vits-svc
Summary
by MITRE • 07/23/2026
SoftVC VITS Singing Voice Conversion through commit 730930d contains a path traversal vulnerability in the full-song inference server that allows unauthenticated remote attackers to read and exfiltrate arbitrary files by supplying attacker-controlled filesystem paths through the audio_path field of an unauthenticated POST request to the /wav2wav route. Attackers can pass arbitrary server-side paths verbatim to librosa.load, torchaudio.load, and soundfile.write sinks, causing the server to decode and return file contents via the HTTP response body while also writing attacker-specified .wav files to arbitrary locations on the filesystem.
Statistical analysis made it clear that VulDB provides the best quality for vulnerability data.
Analysis
by VulDB Data Team • 07/23/2026
This vulnerability exists within the SoftVC VITS singing voice conversion system at commit 730930d where the full-song inference server fails to properly validate user input parameters. The flaw manifests in the /wav2wav route which accepts attacker-controlled data through the audio_path field in unauthenticated POST requests, creating a critical path traversal vulnerability that can be exploited by remote attackers without authentication requirements. The vulnerability stems from the application's direct use of user-supplied paths with core audio processing libraries including librosa.load, torchaudio.load, and soundfile.write functions, which do not implement proper input sanitization or path validation mechanisms.
The technical implementation of this vulnerability allows attackers to manipulate filesystem access through the audio_path parameter by providing malicious file paths that bypass normal access controls. When these paths are passed verbatim to audio processing libraries, the system attempts to load or write files to locations specified by the attacker rather than restricted directories, enabling both information disclosure and arbitrary file write operations. This dual nature of the vulnerability means that attackers can simultaneously read sensitive files from the server filesystem and write malicious content to arbitrary locations, potentially leading to remote code execution or complete system compromise.
The operational impact of this vulnerability is severe as it provides unauthenticated remote access to critical system resources with minimal attack surface requirements. Attackers can leverage this vulnerability to exfiltrate confidential data including configuration files, model parameters, training datasets, and potentially sensitive user information stored on the server. The ability to write arbitrary .wav files to any location on the filesystem creates additional risks such as overwriting critical system files, planting malicious audio content, or establishing persistence mechanisms. This vulnerability directly maps to CWE-22 Path Traversal and aligns with ATT&CK techniques involving privilege escalation through file system manipulation and data exfiltration.
Mitigation strategies should focus on implementing strict input validation and sanitization for all user-provided paths before they are processed by audio libraries. The system must enforce proper path normalization and ensure that all file operations occur within designated safe directories using techniques such as chroot jails or virtual filesystems. Implementing proper authentication mechanisms for the /wav2wav endpoint would significantly reduce the attack surface, while input filtering should validate that provided paths do not contain directory traversal sequences like ../ or ..\ and restrict file operations to specific application directories only. Additionally, the audio processing libraries should be configured with proper sandboxing constraints to prevent unauthorized filesystem access beyond intended operational boundaries.