This is the organisation-wide default. A repository that ships its own SECURITY.md — vibey does, because its threat model is specific to running autonomous engines against a working tree — overrides this one, and that copy is what its Security tab shows.

Do not open a public issue. Use GitHub's private reporting on the affected repository: Security → Report a vulnerability. If that is unavailable, email adam@matthewsteinberger.com with SECURITY in the subject.

Please include: the repository and version, what an attacker gains, the smallest reproduction you have, and whether it is already public.

What to expect: acknowledgement within 3 working days; an assessment with a severity and a rough timeline within 10. You will be credited in the release notes unless you ask not to be. This is a small project maintained by one person — there is no bounty, and no SLA beyond making a genuine effort.

Please give a reasonable window to ship a fix before disclosing publicly.

Only the latest released version of each package receives security fixes. These are pre-1.0 projects on a fast release cadence; there are no long-term support branches. Pin a version, and upgrade to take a fix.

• The published packages: vibey, vibey-gh, vibey-skills, vibey-bootstrap, and the *loop runners.
• The GitHub Actions workflows these projects render into a repository that adopts them — in particular anything reachable from a pull_request_target trigger, which runs with the base repository's permissions.
• Secret handling: anything that could cause a token, key or credential to be written to a log, an artifact, a published page, or a commit.

• Vulnerabilities in the autonomous engines themselves (Claude Code, Codex, Cursor, Antigravity, Qwen) — report those to their vendors.
• Anything that requires an attacker to already have write access to the repository or to the machine. These tools execute code from the working tree by design; a contributor who can change that tree can run code, and that is a property of the tool rather than a flaw in it.
• Findings from a scanner with no demonstrated impact.

Execution is the product. These projects run agents that edit and execute code in a working tree. Run them against repositories you trust, on a machine you are willing to let them change. Isolation is per-worktree by default.

Provenance is enforced, not advisory. Every commit carries an attribution trailer and every source file a header, checked in CI. That is an integrity and authorship control — it tells you what produced a change. It is not an access control, and it does not authenticate anyone.