Claude Security Skills

Security skills for Claude Code. Install them once and ask Claude, in plain language, to scan a repo for leaked secrets, review Python code, red-team an LLM for prompt injection, or audit HTTP headers, JWTs, Dockerfiles, CORS, and dependencies. Claude picks the right skill, runs it, and explains what it found.

Everything here runs on the Python standard library — no packages to install, nothing phoning home. The analysis runs offline; only the few skills that need to hit a URL use the network, and only when you ask them to.

Other languages: Español · Русский

The skills

Skill	What it does	Engine
secret-scanner	Finds hardcoded API keys, tokens and private keys using vendor patterns plus Shannon-entropy analysis, tuned for few false positives	Custom entropy engine
sast-lite	AST-based static analysis for Python: command injection, eval/exec, insecure deserialization, SQLi, weak crypto, disabled TLS — each tagged with a CWE	Python ast walker
prompt-injection-tester	Red-teams your own LLM app with a categorized payload library and canary detection, then scores resilience 0–100	Canary harness
http-sec-audit	Checks HTTP security headers and cookie flags (CSP, HSTS, SameSite, …) and gives concrete fixes	urllib + pure core
jwt-inspector	Decodes and audits JWTs (alg=none, weak expiry, claim hygiene) and cracks weak HMAC secrets offline	HMAC + checks
dependency-check	Flags known-vulnerable and unpinned deps in requirements.txt, package.json and pyproject.toml; offline DB plus optional OSV.dev	Version matcher
dockerfile-scan	Catches insecure Dockerfile patterns: running as root, :latest base images, curl | sh, remote ADD, baked-in secrets	Dockerfile parser
cors-auditor	Audits CORS config for wildcard-with-credentials, reflected origins, null origin and overly broad methods

Each skill is self-contained, has its own tests, and exits non-zero when it finds something — so it also works as a CI step.

Install

Copy the skills into your project's .claude/skills/ directory:

git clone https://github.com/NovaCode37/claude-security-skills.git
cp -r claude-security-skills/skills/* .claude/skills/

Or into ~/.claude/skills/ to have them in every project. Restart Claude Code and it discovers them from each SKILL.md. There's nothing else to install.

Usage

Just ask Claude. For example:

You say	Claude runs
"Any secrets committed in here?"	secret-scanner
"Security-review this Python file."	sast-lite
"Is my AI assistant jailbreakable?"	prompt-injection-tester
"Check example.com's security headers."	http-sec-audit
"Decode and audit this JWT."	jwt-inspector
"Are my dependencies vulnerable?"	dependency-check
"Review my Dockerfile."	dockerfile-scan
"Is my API's CORS safe?"	cors-auditor

Every engine also runs on its own from the command line:

python skills/secret-scanner/engine.py .            --json
python skills/sast-lite/analyzer.py src/            --min-severity high
python skills/prompt-injection-tester/attacker.py   --demo
python skills/http-sec-audit/audit.py https://example.com
python skills/jwt-inspector/inspector.py "<token>"
python skills/dependency-check/checker.py requirements.txt
python skills/dockerfile-scan/scanner.py Dockerfile
python skills/cors-auditor/auditor.py https://api.example.com

Here's what a run looks like:

$ python skills/secret-scanner/engine.py .
[secret-scanner] 2 potential secret(s) found:

  CRITICAL   src/config.py:14:18
             Stripe secret key [stripe-secret]  value=sk_l...k1L2 (len=32)
  HIGH       src/config.py:12:11
             AWS Access Key ID [aws-access-key-id]  value=AKIA...MPLE (len=20)

Summary: critical=1, high=1

Tests

pip install pytest
pytest skills/ -q

158 tests, all offline, run in under a second.

Design principles

• No runtime dependencies. Pure Python 3.9+ stdlib, so the skills run in locked-down CI and are easy to read and audit.
• Offline by default. The analysis logic takes data in and returns findings; network access is optional and explicit.
• Few false positives. Entropy thresholds, keyword anchoring and placeholder allowlists keep the noise down.
• CI-friendly. Consistent exit codes (0 clean, 1 findings, 2 error) and --json on every skill.
• Safe by default. Secrets are redacted in output, and the offensive skills are meant for systems you own or are allowed to test.

Contributing

New skills and rules are welcome. The good first issues in docs/GOOD_FIRST_ISSUES.md each say which file to edit and how to know you're done, and CONTRIBUTING.md has the skill template and conventions. If you have an idea, open a discussion.

Legal

These tools are for authorized security testing, learning and defensive work. Only scan systems and data you own or have permission to test. The maintainers aren't responsible for misuse.

License

MIT