Security fixes are applied to the latest release of VoxLIT on the main branch.
| Version | Supported |
|---|---|
| 1.0.x | ✅ |
| < 1.0 (LIT for Voice) | ❌ |
We take the security of VoxLIT seriously. If you believe you have found a vulnerability, please follow these steps:
-
Do not disclose publicly. Do not open a public issue, pull request or discussion about the vulnerability until it has been fixed.
-
Report it privately. Use GitHub's private reporting (Security → Report a vulnerability) on this repository, or contact one of the maintainers listed under Authors directly.
-
Provide details. Please include:
- A description of the vulnerability
- The affected component (Frontend, Backend API, a specific task workbench, or the Modal deployment)
- Steps to reproduce the issue
- The potential impact
- Any suggested fix (optional)
-
Response. We aim to acknowledge reports within 72 hours. VoxLIT is maintained by a small academic team, so fixes are made on a best-effort basis, prioritised by severity.
-
Disclosure. We will work with you to validate the issue, agree on a fix, and credit you once it is released (if you wish).
In scope:
- The FastAPI backend (
Backend/app/), including file upload, inference and analysis endpoints - The React frontend (
Frontend/) - The deployment configuration (
deploy/)
Out of scope:
- Vulnerabilities in third-party models, datasets or dependencies themselves (please report those upstream; we will update our pinned versions)
- Attacks that require access to the Modal account, the Hugging Face token or the host machine
When developing or deploying VoxLIT:
-
Secrets. Never commit tokens or credentials. Keep the Hugging Face token in
.envlocally (already ignored by git) and in thevoxlit-hfModal Secret in production. -
Model checkpoints. Load checkpoints only from trusted, pinned sources. Prefer
torch.load(..., weights_only=True); a checkpoint loaded with full pickling can execute arbitrary code. -
Uploaded audio. Treat uploaded files as untrusted input. Validate file type and size, and do not keep user uploads longer than needed.
-
Data protection. Voice recordings are personal data. Only process audio you have consent to use, and follow the licence terms of each reference dataset (VoxCeleb, AMI, RAVDESS, Common Voice, ASVspoof 2019, ASVspoof 5, In-the-Wild).
-
API exposure. Restrict CORS origins (
ALLOWED_ORIGINS) to the frontend's domain, serve everything over HTTPS, and add authentication and rate limiting before exposing compute-heavy endpoints publicly. -
Dependencies. Keep
Frontend/package.jsonandBackend/requirements.txtup to date, and reviewnpm auditandpip-auditoutput periodically.
Thank you for helping to keep VoxLIT secure!