MLForge handles local datasets and executable Python model artifacts. Please report security problems privately when possible so maintainers can assess them before exploit details become public.

MLForge remains in the 0.y.z development series. Security fixes are made on the latest stable 0.6.x line. Older minor lines and locally modified copies are not supported; unreleased source snapshots receive best-effort review only.

Use the repository's Security tab and Report a vulnerability when that private-reporting option is available. Include:

• the affected MLForge version or commit;
• the operating system and Python version;
• the affected command or Python API;
• minimal reproduction steps and expected impact;
• whether artifact loading or untrusted input is involved; and
• any mitigation you have already verified.

If private reporting is unavailable, open a minimal GitHub issue asking the maintainers to provide a private contact channel. Do not include exploit code, malicious artifacts, credentials, private data, or details that would enable abuse in that public issue.

Maintainers will assess the report, request any information needed to reproduce it, coordinate a fix when warranted, and discuss disclosure timing with the reporter when contact is possible. Response and remediation times are not guaranteed while the project is in its 0.y.z development series.

• A .mlforge artifact contains pickle data and can execute code when loaded. Inspection and checksums do not make an untrusted artifact safe.
• trusted=True and --trust-artifact are explicit source-trust assertions, not sandboxing or authentication.
• MLForge is a single-user local tool. It does not provide authorization, tenant isolation, network serving, artifact signing, or hostile-code isolation.
• Dataset and prediction loaders validate local CSV structure and size, but callers remain responsible for filesystem permissions and sensitive-data handling.
• Profiles and run manifests contain paths, schema/target metadata, metrics, warnings, and bounded error text. Review them before public sharing; they are not anonymized reports.

Read the complete artifact security model before loading, sharing, or deploying a model artifact.