| Version | Supported |
|---|---|
| 1.0.x | Yes |
| < 1.0 | No |
Please do not open a public issue for a security problem.
Report it privately through GitHub: Report a vulnerability
You can expect an acknowledgement within 7 days and, where a fix is needed, a patched release once it is ready. Credit is given in the advisory unless you ask otherwise.
In scope:
- Arbitrary code execution when loading a checkpoint or ONNX model
- Path traversal or arbitrary file write in the dataset or scrape paths
- Command injection through CLI arguments
- Anything in the published site under
docs/
Out of scope:
- Model accuracy, false positives and false negatives. Rakshak is a research preview and its predictions are expected to be wrong sometimes — see the disclaimer in the README.
- Vulnerabilities in PyTorch, torchvision or ONNX Runtime. Report those upstream; tell us if a workaround is needed here.
- The copyright status of the bundled dataset images. That is a known and documented issue — see DATASET.md.
Checkpoints are loaded with weights_only=True. torch.load can execute
arbitrary code when unpickling; Rakshak disables that path, so a downloaded
.pt file cannot run code on your machine through the normal load route. Even
so, prefer checkpoints you trained or that came from a source you trust.
The web demo runs entirely client-side. Images selected in the browser demo are decoded to a canvas and fed to ONNX Runtime in the same tab. Nothing is uploaded and no request carries image data.