SUPREME is published to PyPI as supreme-unlearning. Security fixes are applied to the latest released version on the main branch. Please upgrade to the most recent release before reporting an issue.

Please do not open a public issue for security problems.

Report privately through GitHub's private vulnerability reporting (Security tab -> "Report a vulnerability").

When reporting, please include:

• a description of the issue and its potential impact,
• the exact command or supreme.run_* call that triggers it,
• steps to reproduce, and
• your environment (OS, install method, Python/PyTorch versions).

You can expect an initial acknowledgement within a few days. Once a fix is ready, it will be released and the advisory published with credit to the reporter unless anonymity is requested.

SUPREME pins its full dependency stack (PyTorch 2.1.0, Lightning 2.1.0, Transformers 4.35.0, DeepSpeed 0.14.5, ...) to the exact environment used to produce the results in the WIPE-OUT 2 (ECML-PKDD 2026) paper. These versions are interdependent (e.g. DeepSpeed 0.14.x is the last line compatible with PyTorch 2.1), so they are kept frozen for reproducibility rather than upgraded as new dependency advisories appear.

Published advisories against these pinned versions (e.g. torch.load / checkpoint deserialization issues) are exploitable only when loading untrusted checkpoints, weights, or Hub models. SUPREME is a research framework intended to train, unlearn, and evaluate models from sources you control:

• Only load checkpoints and datasets you produced yourself or obtained from a source you trust.
• Do not point SUPREME at model files or Hub repositories from unknown parties.
• Run experiments in an isolated environment (venv/container) as an unprivileged user.

Dependabot alerts covered by this policy are dismissed as "tolerable risk" with a pointer to this document. Vulnerabilities in SUPREME's own code are always in scope - please report them as described above.