Adduce is beta software. Security fixes are made on the active development branch and released in the next stable version. Routine backports are not provided.

Users should upgrade to the latest stable release before reporting an issue that may already have been corrected:

python -m pip install --upgrade adduce

Please do not open a public issue for a vulnerability or include credentials, private repository content, unpublished results, or personal data in a report.

1. If the repository's Security page offers Report a vulnerability, use GitHub's private advisory form. Availability depends on the repository's current GitHub configuration.
2. If that form is unavailable, email chudasama.h@northeastern.edu with the subject Adduce security report.

Include the affected Adduce version, operating system, command or mode, a minimal reproduction, the expected security boundary, the observed behavior, and the likely impact. Redact all secrets and sensitive research data. If a test repository is needed, use a minimal synthetic example that you are authorized to share.

Test only repositories and systems you own or are authorized to assess. Stop after demonstrating the issue, avoid disrupting third-party services, and do not retain or disclose data that is not yours.

The maintainer will coordinate validation, remediation, and disclosure with the reporter. No fixed response or release deadline is promised for this beta project.

The default built-in static audit does not execute target repository code or initiate remote-metadata requests. Several explicit features cross additional trust boundaries:

• adduce check --online and adduce pin-remotes --diff make outbound HTTPS requests for repository-derived artifact identifiers.
• adduce reproduce --yes executes the selected repository command twice. Repository copying provides input isolation only; it is not a process, credential, filesystem, device, resource, or network sandbox.
• adduce-rng-audit --yes ... imports RNG libraries and executes the selected script unsandboxed in the Adduce process. Its warning and confirmation gate do not provide containment.
• Installed rule and reporter plugins are Python code and must be trusted as part of the Adduce environment.
• adduce checklist --llm sends the documented evidence summary to the provider explicitly configured by the user.

Likely-credential findings report a location and kind without the matched value. This is a narrow finding-level control, not output sanitization or data loss prevention. Generated drafts can include repository-derived commands, paths, identifiers, and metadata. .adduce/reproduce-report.json records the selected command and parsed numeric metric names and values verbatim. Do not place credentials in repository commands or metric names, and inspect every generated file before sharing it.

The complete boundary, implemented controls, residual risks, and recommended operating practices are documented in the security model.

False positives, false negatives, scoring disagreements, and documentation errors are important but can normally be reported through the public issue tracker when the report contains no sensitive information. A repository command doing something harmful after the user explicitly starts adduce reproduce --yes is not, by itself, a sandbox escape because that command is not sandboxed. Report it privately if Adduce executed the command without consent, exposed data contrary to the documented boundary, or allowed the command to bypass an implemented containment control.