Audit Code

Security-focused code review for hardcoded secrets, dangerous calls, and common vulnerabilities

Security Scan

The skill is internally coherent for a repository security auditor: it contains a local scanner, requests no credentials, has no install steps, and its behavior matches its description except for a small documentation-versus-code mismatch.

Purpose & Capability

Name/description (code security review for secrets, dangerous calls, dependencies, permissions) align with included artifacts: a Python scanner (scripts/audit_code.py) and a pattern database (scripts/patterns.py). No unrelated credentials, binaries, or external services are requested.

Instruction Scope

Runtime instructions tell the agent to run the included Python script against a target path and to use Bash/Read/Glob/Grep tools. The scanner legitimately reads repository files (including .env and key files) and checks permissions; this is expected. Minor inconsistency: SKILL.md claims 'If $ARGUMENTS is empty, default to $PROJECT_ROOT', but the script actually requires an explicit path and exits if none is provided. Also the frontmatter pre-approves Bash (allowed-tools includes Bash), which is functional here but is a capability that should be intentionally allowed.

Install Mechanism

No install spec or remote downloads; the skill is instruction+bundled code only. All code is included in the package (no external fetches or extract operations), which reduces install-time risk.

Credentials

The skill declares no required environment variables or credentials. The scanner will read repository files (including potential secret files) as part of its function; that file access is proportionate to an auditing tool and is not requesting unrelated secrets.

Persistence & Privilege

always is false and disable-model-invocation is true (skill cannot be auto-invoked by the model), and it does not request persistent system-wide changes. The only elevated capability in the frontmatter is allowing Bash execution when run; because model auto-invocation is disabled, the skill can only be run explicitly by a user.

Assessment

This skill appears to be a legitimate local code auditor. Before running it: (1) review the bundled scripts yourself (they are included) to confirm behavior; (2) run it against a specific project path (the script requires a path — SKILL.md's claimed default is inaccurate); (3) run it as a non‑privileged user or inside a sandbox/container to avoid accidental scanning outside the intended repository; (4) be aware it will read files like .env and private keys (that is its purpose) but it does not contain obvious network/exfiltration code; and (5) if you plan to let other agents/tools invoke it, consider the implications of allowing Bash execution in the frontmatter even though model auto-invocation is disabled.

audit-code -- Project Code Security Review

Security-focused code review of project source code. Covers OWASP-style vulnerabilities, hardcoded secrets, dangerous function calls, and patterns relevant to AI-assisted development.

What to do

Run the auditor against the target path:

python3 "$SKILL_DIR/scripts/audit_code.py" "$ARGUMENTS"

If $ARGUMENTS is empty, default to $PROJECT_ROOT.

What it checks

• Hardcoded secrets -- API keys (AWS, GitHub, Stripe, OpenAI, Slack), tokens, private keys, connection strings, passwords
• Dangerous function calls -- eval, exec, subprocess with shell=True, child_process.exec, pickle deserialization, system(), gets(), etc.
• SQL injection -- String concatenation/interpolation in SQL queries
• Dependency risks -- Known hallucinated package names, unverified installations
• Sensitive files -- .env files committed to git, credential files in repo
• File permissions -- Overly permissive chmod patterns
• Exfiltration patterns -- Base64 encode + network send, DNS exfiltration, credential file reads

Output

Structured report with severity-ranked findings, file locations, and actionable remediation steps.

When to use

• Before committing or pushing code
• When reviewing third-party contributions or PRs
• As part of a periodic security audit of the codebase
• After AI-assisted code generation to verify no secrets or vulnerabilities were introduced

Advisory hooks

The repository's .claude/settings.json includes PreToolUse hooks that warn on dangerous Bash and Write operations. These hooks are advisory only -- they produce warnings but do not block execution.

• audit-code is the detection layer for source code security issues
• The hooks provide supplementary runtime warnings during agent operation
• To enforce blocking, hooks must return {"decision": "block"} instead of warning messages