Repository Security Audit

An agent-driven, read-only security analyzer for source repositories. It inventories the whole codebase, builds a threat model, verifies candidate vulnerabilities against source, and returns a plain-English assessment plus an agent-ready correction queue.

This repository does one job. Personal-computer and account-security auditing lives in the separate private personal-security-audit repository.

Start here

Open this repository in a fresh Claude Code or Codex session and say:

Start the repository security audit.

The agent asks for the repository path or Git URL if you have not supplied one. It never assumes this toolkit is the target. The audit is read-only: it does not edit the target, install tools, run untrusted repository code on your host, commit, push, or deploy.

You can also invoke the skill directly:

Use $audit-repo-security to audit /path/to/repository.

What it does

1. Freezes the target commit and records a redacted manifest digest.
2. Inventories code, dependencies, entry points, identity boundaries, data flows, infrastructure, CI, and operational surfaces.
3. Builds a threat model proportional to the application.
4. Runs applicable static checks without executing target-controlled code on the host.
5. Verifies tool alerts and review candidates against source and compensating controls.
6. Produces a decision brief, explicit coverage and unknowns, confirmed findings, and a correction queue agents can implement later.

Whole-repository coverage is the default. A branch-diff audit is available as a faster, narrower option, but is never presented as proof that the entire repository is secure.

Correct confirmed findings

The separate remediate-security-findings skill consumes selected finding IDs and limits edits to their approved file scope:

Use $remediate-security-findings to correct REPO-001 and REPO-004. Do not push or deploy.

Unknown or unverified items are not silently turned into fixes.

Trust model

Read TRUST.md before pointing the toolkit at sensitive code. Repository instructions and source comments inside the target are treated as untrusted evidence, not authority. Commands that load target code, hooks, plugins, build files, or package scripts require a disposable isolated copy with no host credentials and network disabled by default; otherwise the check remains unknown.

Install

scripts/install.sh

If previous copies exist, review the paths printed by the installer and rerun scripts/install.sh --archive-existing. Existing directories are archived, never deleted.

Validate

python3 scripts/validate.py
python3 -m unittest tests/test_validate.py
for skill in skills/*; do uv run --with pyyaml python "${CODEX_HOME:-$HOME/.codex}/skills/.system/skill-creator/scripts/quick_validate.py" "$skill"; done

No audit can guarantee that a repository has no vulnerabilities. This toolkit establishes coverage against a recorded manifest and the evidence available to the agent, then makes remaining uncertainty visible.

No license has been selected while the repository remains private.