Email ops@themalwarefiles.com. Please do not open a public issue.

Include what you did, what happened, and which version you were running. senbonzakura doctor prints the version and the pinned tool versions, and pasting its output saves a round trip.

You will get a reply within seven days. This is a project with one maintainer, so that is a realistic promise rather than a generous one. If a week passes with no reply, send the email again; it is far more likely to have been missed than ignored.

This tool removes a model's refusal behaviour. A model that answers a harmful request after being abliterated is the tool working as documented, not a vulnerability. Please do not report it as one, and please do not include the generation.

What is in scope is anything that harms the person running the tool, or anyone downstream of a number it produced:

• Code execution, file reads or file writes outside the paths you pointed it at, including through a crafted checkpoint, dataset, GGUF or result file. This tool loads other people's files, so this is the category that matters most.
• Credentials, tokens or local paths reaching a log, a result artefact, a commit or the terminal.
• A measurement that can be made to report a number that is wrong in a way the report does not disclose. A tool whose whole claim is that its figures carry their conditions has to treat a silently wrong figure as a security problem, not merely a bug.
• Anything in the published package that is not built from the tagged source.

Confirmed reports get a fix, a CHANGELOG entry describing the problem in plain language, and credit in that entry unless you ask us not to. If a published figure turns out to be affected, the figure is withdrawn in the documentation as well as fixed in the code. That has happened before and it is the policy rather than an exception.

Where a dependency has a known advisory and no fixed release, the advisory is recorded in pyproject.toml under [tool.senbonzakura.security], together with what was done about it and what is still exposed. A test fails if the assessed version moves, so an accepted risk cannot be inherited without being re-read. If you think one of those assessments is wrong, that is a report worth sending.

The most recent release. This project does not backport.