open-sesame is a suite of skills that turns Claude into a disciplined penetration
tester. Point it at a target you're authorized to assess and it runs the whole
engagement — maps the attack surface, works methodically through an industry
methodology, validates findings with the lightest safe proof, and writes clean reports
for both engineers and leadership.

It's built as a set of Markdown instructions and small helper scripts — no model
fine-tuning, no binaries. You can read every line of what it tells Claude to do.

[!WARNING]
Authorized testing only. This skill is for systems you own or have explicit
permission to test (your own apps, a signed client engagement, an in-scope
bug-bounty target, or a CTF / practice lab). It is designed to refuse unauthorized
targets, treat the platform/host a target runs on as out of scope by default, and
prefer non-destructive proofs. Using it against systems you don't have permission to
test may be illegal. You are responsible for staying within scope and the law.

✨ What it does

• Authorization-first — both skills confirm scope and rules of engagement before any
  active test, and treat third-party platform/cloud infrastructure as out of scope.
• Web (web-pentest) — full OWASP WSTG coverage: recon, configuration, auth, session,
  access control (incl. multi-tenant isolation), injection, crypto, business logic,
  client-side, and API tests; safe RCE validation and reverse shells.
• Network (net-pentest) — PTES-based external and internal testing: host discovery,
  port/service scanning, enumeration, vuln ID, exploitation, privilege escalation (Linux +
  Windows), pivoting, and Active Directory attacks (Kerberoasting, BloodHound, NTLM
  relay, the path to Domain Admin) — plus hybrid/cloud identity (Entra ID, AWS),
  Wi-Fi, and container/Kubernetes testing.
• Tooling playbooks — concrete, safe-usage workflows for Burp, ffuf, nuclei, sqlmap,
  nmap, masscan, netexec, impacket, responder, BloodHound, and more.
• Three reports, two audiences — a technical findings report, a plain-language
  stakeholder summary, and a how-to-recreate guide, each rendered to PDF.

🚀 Quick start

1. Install the skill

Clone into a location Claude discovers skills from:

# Per project (available inside one repo) — installs both skills
git clone https://github.com/hassan-algo/open-sesame.git
cp -r open-sesame/.claude/skills/* .claude/skills/

# Or personal (available in every project)
cp -r open-sesame/.claude/skills/* ~/.claude/skills/

2. Use it

In Claude Code, just describe the work — the right skill triggers automatically:

Pentest my web app at https://staging.myapp.example — I own it. Start with recon.
Scan my network 10.0.0.0/24 and enumerate Active Directory — it's my lab.

web-pentest triggers on things like "find vulns in my web app", "test for IDOR/SQLi",
"validate this RCE"; net-pentest on "scan this subnet", "enumerate these hosts",
"Kerberoasting / BloodHound", "privilege escalation on this box". Or invoke explicitly
with /web-pentest or /net-pentest.

3. Get PDF reports

After a report is drafted in Markdown, render it to a styled PDF:

# each skill ships the same renderer
.claude/skills/web-pentest/scripts/render-report.sh report.md report.pdf

The script converts Markdown → HTML (pandoc, else npx marked) → PDF (headless
Chrome/Chromium, else weasyprint, else macOS cupsfilter). Install at least one tool
from each stage.

🗂️ How it's organized

.claude/skills/
├── web-pentest/                 # single web application (OWASP WSTG)
│   ├── SKILL.md                 #   authorization, workflow, pointers
│   ├── references/              #   methodology, tooling, post-exploitation,
│   │                            #   reporting, lab-targets
│   ├── assets/report-template.md
│   ├── scripts/render-report.sh
│   └── evals/                   #   test cases + benchmarks
└── net-pentest/                 # networks & Active Directory (PTES / NIST)
    ├── SKILL.md
    ├── references/              #   methodology, active-directory, privesc,
    │                            #   pivoting, password-attacks, cloud-identity,
    │                            #   wireless, containers, tooling, reporting,
    │                            #   lab-targets
    ├── assets/report-template.md
    ├── scripts/render-report.sh
    └── evals/evals.json         #   trigger/behavior test cases

Supporting folders at the repo root:

• index.html — the project landing page.
• reports/ — generated engagement reports. Keep them local: reports/ is not yet in
  .gitignore, so add it before committing.

Each skill uses progressive disclosure: SKILL.md is always loaded; the references/
files are pulled in only when a phase needs them, keeping the context lean and the
workflow fast.

🧪 Validation

web-pentest was built and tuned with Anthropic's skill-creator and validated across
two eval iterations — methodology, reporting, authorization-gating, RCE/reverse-shell
guidance, cross-tenant IDOR, and scope discipline. Summary results live in
.claude/skills/web-pentest/evals/benchmarks/.

Beyond the evals, see BENCHMARKS.md for timed, end-to-end
engagements, and CASE-STUDY.md for the story behind each one:

Dead ends and lab-injected findings are documented, not hidden.

🤝 Contributing

Contributions are very welcome — this gets better the more practitioners shape it. Good
first contributions:

• Extend the methodology — add or sharpen WSTG test cases in references/methodology.md.
• Add tooling recipes — real-world command workflows in references/tooling.md.
• Improve reporting — templates, severity guidance, or the render-report.sh backends.
• Add eval cases — new prompts in evals/ that catch gaps or regressions.

See CONTRIBUTING.md for how to propose changes. Please keep everything
aligned with the skill's core principle: authorized, non-destructive, methodical
testing — contributions that add offensive capability without that framing won't be
merged.

📜 License

MIT — free to use, modify, and share.

🙏 Acknowledgements

• Methodology based on the OWASP Web Security Testing Guide
  & Top 10, the PTES,
  NIST SP 800-115, and MITRE ATT&CK.
• Built as Claude Agent Skills.