/audit-security

/audit-security audits a whole project for security, not one module. It runs the real scanners (osv-scanner for dependency CVEs, gitleaks for secrets in the working tree and in full git history, semgrep for SAST, trivy for infrastructure-as-code), builds a threat model so findings are ranked by what is actually worth attacking, hunts the project's own silent-failure invariants, then performs its own LLM-driven review of the entire repository and adversarially triages every candidate. It profiles the project before it checks anything, so it is stack-agnostic, and it is report-only unless you pass --fix.

Install

From your project root:

mkdir -p .claude/skills

ln -s ../../node_modules/@daanvandenbergh/claudekit/skills/ts/audit-security \
      .claude/skills/audit-security

Restart Claude Code so it picks up the new skill.

External tools. The skill shells out to real scanners and requires osv-scanner, gitleaks and semgrep on your PATH. trivy is required only when the repo contains IaC markers (*.tf, Kubernetes YAML, a Dockerfile). Phase 00 checks each tool's presence, version and a canary probe, and hard-halts if a required one is missing or is running zero rules - it prints the install command rather than reporting a partial scan as clean.

Run it

/audit-security [scope] [--fix] [--quick] [--no-history] [--out <path>] [--verify-secrets]

Argument	What it does	Default
scope	A path, area or subsystem that narrows the Phase 40 deep review. The scanners, the secret scan and the invariant hunt still cover the whole repo.	Whole project
--fix	Opt into the tiered remediation phase. Everything before it is read-only.	Off (report-only)
--quick	Runs phases 00, 10, 20, 30, 50 and 60 only - scanners, invariants, triage, report. Skips the LLM sweep and the ASVS net, and the report is stamped so it never reads as a full run.	Off
--deep	Adds the noisier semgrep packs (p/security-audit, p/default) to the SAST sweep.	Off
--since <ref>	Deep-audits only code changed since <ref> plus its reverse-dependency closure. Every unfixed High+ finding from the prior ledger is carried forward.	Off; <ref> defaults to the last run's commit
--resume	Resumes an interrupted run from .agentstore/audit-security/state/.	Off
--budget <tier|tokens>	Caps spend and stamps the depth tier reached (T0/T1/T2) on the verdict. Below the always-runs floor the run halts as INCOMPLETE.	Off (no cap)
--no-history	Skips the gitleaks full-history secret pass. The working-tree scan still runs.	Off (history is scanned)
--verify-secrets	Adds TruffleHog live-verification of a candidate secret. This authenticates to the vendor with the credential, so it is an explicit opt-in.	Off
--out <path>	Persists the report to <path>, creating parent directories.	Off (report is inline)
--self-test	Maintainer-only: runs the calibration harness against fixtures/ and emits a precision/recall scorecard instead of auditing.	Off

What it runs

The skill reads the code on disk - your working tree, uncommitted edits included - so you can fix something and re-run before committing. The one exception is the secret scan, which also covers full git history, because a committed-then-deleted credential is still live in every clone.

The phases, in order:

• 00 profile + tooling gate - learns the stack, entrypoints, auth mechanism, tenant key and money representation, then proves every required scanner actually runs.
• 05 scope, 10 threat model - risk appetite, exclusions, and the register of attack surfaces.
• 20 scanners - osv-scanner (scan source -r), gitleaks in both dir and git mode, semgrep with p/ci + p/owasp-top-ten + p/secrets + the detected language pack, and trivy when IaC is present. Each dependency CVE gets a reachability verdict, which downgrades but never auto-suppresses.
• 30 invariants - hunts the project's own rules from .agentstore/audit-security/rules.md. If that file is missing, this phase offers to bootstrap it from past audit reports, git history and the test suite.
• 40 LLM whole-repo review - a local sweep of all the code, directed by the threat model. Skipped by --quick.
• 50 triage - a three-role panel adjudicates each candidate. The ledger is append-only: a rejected candidate stays visible in a Refuted section, and "I could not build a trace" is reported at full severity rather than dismissed.
• 55 compliance (only when a regime is set), 60 report.
• 70 fix - runs only with --fix.

It is report-only by default. With --fix, remediation is tiered by how dangerous the fix is, not how bad the hole is: safe changes (a dependency bump to its minimum fixed version, pinning an action to a SHA) apply automatically and are re-proved against the build and test suite; behavioural fail-closed changes are shown as a diff and applied only on approval; auth, crypto, tenant-scoping, payment paths and anything touching a leaked credential are human-only and produce a runbook, never a patch. A leaked secret is never closed by deleting the line - it stays rotation-pending until you rotate it at the provider.

A report with no blocking findings is not a certificate that the project is secure, and the skill says so in every report.

Its companion rule

rules/audit_security_rules.md is the CLAUDE.md snippet that pairs with this skill. It makes maintaining .agentstore/audit-security/rules.md - your project's own silent-failure invariants, the ones no generic scanner will ever find - a standing obligation rather than a one-time act, and it specifies the format Phase 30 expects: a failure mode, a ✗/✓ pair drawn from a real bug, and a hunt block with a count floor.

Import it like any other rule; see Rules. Run state lives in .agentstore/audit-security/state/ (gitignore it) and accepted risks in .agentstore/audit-security/accepted.md.

When to use it

Use /audit-security when you want the whole project checked for security: before a release, after an auth, money, tenant or AI change, or when you need a dependency and secret sweep. Use /audit instead when you are hardening one module, /audit-deep-logic for cross-module logic bugs that are not security, and /audit-tests for test coverage.

Examples

/audit-security

/audit-security src/billing --out .agentstore/audit-security/report.md

/audit-security --quick --no-history

Related

• Skills and rules - how the symlink and @-import mechanisms work.
• Getting started - installing the package and linking your first skill.