difflab is a research codebase. Only the main branch receives fixes; there are no maintained release branches.

Report privately through GitHub's security advisory form rather than opening a public issue. Expect an acknowledgement within seven days.

Please include the affected file or entry point, the steps to reproduce, and what an attacker gains. A proof of concept helps but is not required.

In scope:

• Code execution or file writes triggered by the config loader (src/difflab/config.py) parsing an untrusted YAML file.
• Path traversal in the checkpoint and dataset loaders.
• Command injection in the CLI (src/difflab/cli.py) or the scripts under scripts/.
• Vulnerable dependency versions that difflab pins directly.

Out of scope:

• Vulnerabilities in PyTorch, diffusers, or other upstream dependencies — report those to their maintainers.
• The quality, bias, or content of generated samples.

difflab.hub and the --resume paths load checkpoints through PyTorch and the Hugging Face Hub. A checkpoint is executable data: loading one from an untrusted source can run arbitrary code. Only load weights you produced or that come from a publisher you trust, and prefer safetensors where the upstream repository offers it.

This is a property of the model formats themselves, not a difflab bug, so reports amounting to "loading a malicious checkpoint runs code" will be closed as out of scope.