Security fixes are applied to the latest released version and the main branch. Pre-release branches may change without a compatibility guarantee.

Do not open a public issue for a vulnerability, exposed credential, private dataset, or personally identifying record. Use GitHub's private vulnerability reporting for this repository. Include only the minimum synthetic information needed to reproduce the problem.

If a credential has already been exposed, revoke or rotate it with its provider before changing local files. Do not paste the credential into a report, commit, chat, screenshot, or Regression Card. Rewriting Git history is a separate, potentially disruptive action and should be planned only after rotation and current-file redaction.

Please allow the maintainer a reasonable period to reproduce and address a report before public disclosure. There is no guaranteed response or remediation time. Good-faith reports that avoid privacy violations and service disruption are welcome.

SFTGuard is designed for local analysis. The command-line tool must not upload datasets, prompts, responses, model weights, or generated reports. Its public static viewer has no report-upload endpoint, analytics, or server-side report processing; a report selected in the viewer remains in that browser context.

Generated artifacts are intended to contain aggregates, configuration, content-addressing information, and bounded findings rather than raw dataset text. This is a design boundary, not a guarantee that every user-supplied file is safe to publish. Before sharing any artifact, follow docs/SHARING.md. Treat hashes as stable identifiers that can still reveal equality or permit guessing when the underlying data is simple.

When reporting a suspected privacy failure:

• name the affected command, version, and output field;
• describe the data category, not the sensitive value;
• use synthetic reproduction data whenever possible;
• remove local usernames, absolute paths, access tokens, and model credentials;
• do not attach the original dataset or an unsanitized report.

Reports about unsafe parsing, path traversal, command construction, secret or raw-text leakage, unbounded resource use, cross-site scripting in the viewer, and dependency compromise are in scope. A PASS Regression Card is not a safety certification and does not show that a model, adapter, or dataset is secure.