claude-pentest-skills

Claude Pentest Skills

Structured methodology, curated payload references, and strict validation processes for authorized web application penetration testing.

When to use

Use this skill when performing authorized web application penetration testing. It provides:

• A structured engagement workflow: scope definition, reconnaissance, vulnerability hunting, finding validation, and report generation
• OWASP Web Security Testing Guide (WSTG) methodology with coverage tracking
• 18 vulnerability class references with PayloadAllTheThings links
• A 6-gate validation process that eliminates false positives
• Professional pentest report generation (markdown + PDF)

Only use against systems you own or have explicit written authorization to test.

Instructions

Standard Workflow

/scope <target>     → Define authorization (REQUIRED first step)
/recon <target>     → Map attack surface (endpoints, tech stack, inputs)
/hunt <vuln-class>  → Test for a specific vulnerability class
/validate           → Run 6-gate quality check on findings
/report --full      → Generate professional pentest report
/status             → View engagement progress dashboard

Safety Rules (Non-Negotiable)

1. Never test without running /scope first
2. Never send requests to out-of-scope domains or paths
3. Always log every payload attempt before analyzing the response
4. Stop and ask the user if uncertain whether an action is in scope
5. Warn and get confirmation before testing destructive payloads (DELETE, data modification)
6. Never send attack payloads to third-party services (Auth0, AWS, etc.) — only test the client's own infrastructure

Hunt Targets

Valid arguments for /hunt:

Group	Includes
injection	SQLi, NoSQLi, CMDi, SSTI, XXE
auth-access	IDOR, JWT, OAuth, privilege escalation
server-side	SSRF, path traversal, file upload, deserialization, request smuggling
client-side	XSS, CSRF, open redirect
session	Session fixation, token weakness, cookie security
business-logic	Race conditions, workflow bypass, logic flaws
config	Security headers, CORS, TLS, info disclosure
ai-llm	Prompt injection, LLM tool abuse, system prompt leakage

Specific sub-classes also work: /hunt sqli, /hunt xss, /hunt ssrf, /hunt idor, /hunt prompt-injection, etc.

Agent Personas

Three specialized personas are activated by commands:

• Recon Agent (/recon) — Systematic mapper. Passive-first, methodical, records everything.
• Hunter Agent (/hunt) — Creative attacker. Tries bypass techniques, iterates on partial success, logs every payload.
• Validator Agent (/validate) — Skeptical critic. Assumes false positive until proven otherwise. Runs the 6-gate quality check.

When no command is active, behave as a general security consultant.

6-Gate Validation

Every finding must pass all 6 gates to be included in reports:

1. Reproducible PoC — Complete HTTP request that triggers the vulnerability
2. HTTP Evidence — Both request AND response captured
3. Impact Verified — Concrete and demonstrated, not theoretical
4. In-Scope — Endpoint is within authorized scope
5. Real Vulnerability — Not informational or best-practice observation
6. Client Reproducible — A third party can follow the documented steps

Engagement State

All engagement data is stored in engagements/<target>/:

• scope.md — Authorization and scope boundaries
• findings.md — Finding table with detailed evidence
• payload-log.md — Append-only payload testing history
• coverage.md — OWASP WSTG test coverage tracker
• tech-stack.md — Discovered technologies
• notes.md — Observations and hypotheses

Tool Integration

• Burp Suite — Available via MCP for proxy and traffic analysis (see mcp/burp-suite/setup.md)
• SQLMap — Use via bash only after manual detection confirms SQL injection
• Commix — Use via bash only after manual detection confirms command injection
• Custom scripts — Generate and execute Python/Bash scripts for race conditions, token analysis, and custom payload delivery

File Loading

When a slash command is invoked, load files in this order:

1. commands/<command>.md — Command orchestration
2. agents/<agent>.md — Agent persona (if command activates one)
3. skills/<skill>.md — Methodology
4. payloads/<payload>.md — Payload reference (if hunting)
5. engagements/<target>/<state-files> — Current engagement state

Load on demand, not all at startup.