src-hunter

A Claude Code skill for SRC / bug-bounty / pentest workflows. Loads automatically when you mention things like bug bounty, hackerone, waf bypass, or 任意账号, and walks Claude through a five-phase hunt: intake → recon → enum → hunt → report.

It bundles a knowledge base built from public sources: nineteen attack-class playbooks, three hundred-odd structured payloads with WAF/EDR bypass variants, the full HackerOne hacktivity feed (High and Critical, disclosed) at the time of build, and statistical residue from the old WooYun corpus.

Install

Marketplace:

/plugin marketplace add MyuriKanao/src-hunter-skill
/plugin install src-hunter@src-hunter

Plain git:

git clone https://github.com/MyuriKanao/src-hunter-skill.git ~/.claude/skills/src-hunter

What's in here

references/
  methodology/    five-phase workflow, attack-priority, bypass toolkit, evidence rules
  playbooks/      one file per attack class — each ends with real H1 cases and a payload library
  industry/       banking/finance and telecom/ISP verticals
  dictionaries/   default credentials and component fingerprints (Chinese stack)
  templates/      submission template (CVSS 4.0)
  h1-reports/     2887 disclosed reports, raw + grouped by weakness
  payloader/      305 structured payloads, 263 WAF/EDR bypass steps, 114 tool cheats

The playbooks are the entry point. They're written from a black-box hunter's perspective — assume you only have a URL — and each one carries the same shape: where to look, what to throw, what response shapes to watch for, how to escalate, and the legal lines.

MCP integration

The skill wires in a local MCP server as a tool layer so Claude can drive browser automation, CDP debugging, network interception, JS hooks, AST deobfuscation, Frida memory verification, WASM reversing, source-map reconstruction, Android adb bridging, and SSL-pinning bypass directly during the hunt phase.

Current pick: jshookmcp 0.3.0 (134 curated / 386 full / 36 domains). The full index and scenario→tool map lives in references/tools/mcp-jshook.md.

Seven high-affinity playbooks (xss / rce / ssrf-cache-host / mobile / oauth-saml-jwt / api-rest / file-upload) carry a ## 相关 MCP 工具 trailer block telling Claude which jshook tools to reach for and when.

TODO

Support more MCP tools
Multi-agent execution workflow

Triggers

The skill loads on any of these (and a few more spelled out in SKILL.md):

bug bounty, hackerone, src 挖洞, 漏洞赏金, 众测
waf bypass / 绕过 WAF
"how do I test this endpoint / API / parameter"
任意账号 / 任意修改 / 任意删除, 密码重置 / 找回密码
默认凭据, Actuator, exposed admin, etc.

Or invoke explicitly: /src-hunter <target>.

Playbook list

Playbook	H1 cases bundled
arbitrary-x-authz (IDOR / account takeover / privesc)	465
rce (deserialization / SSTI / XXE / framework)	385
xss	335
info-disclosure	319
oauth-saml-jwt	240
logic-flaws (CSRF, clickjacking, payment)	234
path-traversal / LFI / RFI	163
sqli	147
dos	138
ssrf-cache-host	108
unauth-access (default creds, Actuator, exposed services)	46
http-smuggling / CRLF	38
api-rest / WebSocket	15
file-upload	8
mobile (Android / iOS)	8
race-conditions	5
llm-prompt-injection	1
graphql	1
intranet-postexp (post-exploitation reference)	—

Sources

HackerOne hacktivity feed — 2887 disclosed High/Critical reports, fetched directly. Public data.
WooYun historical archive — 88,636 cases. Only parameter-frequency tables, case IDs, and bypass patterns are retained. The platform is defunct; data is unrecoverable elsewhere.
Payloader — 305 structured payloads, 263 WAF/EDR bypass steps, 114 tool cheats from the open-source 3516634930/Payloader repo.

This project re-organizes and translates these sources into a hunter-oriented skill. No proprietary data, no scraping of authenticated content.

Rules of engagement

Each playbook spells out concrete limits. The most-commonly-stepped-on ones:

Sample size: for SQLi, proving the database name / version is enough — don't dump rows. For IDOR / Mongo / ES exfil, 1–3 samples is the cap.
Self-account testing: authz bypass, password reset, JWT forgery, redirect_uri abuse, blind XSS — all on two accounts you registered yourself. Never on a stranger's account, even if you can.
Read, don't write: with RCE, only id / whoami / uname -a. With unauth Redis / Mongo, only info / ping / db.version(). With LFI, stop at root:x: — don't read /etc/shadow.
No real side effects: no real SMS, no real charges, no real emails, no real refunds, no overwriting files, no editing global announcements / mail templates. Stop once the endpoint returns 200.
DoS / concurrency: each repro ≤ 60 s, serial, 5 reps total. Race conditions: 50–100 concurrent, never 1000+. Rate-limit bugs: 5–10 hits to your own number, not 1000.
Don't keep loot: webshells, heapdumps, backups, dumped source — local only, deleted once the report lands. Never push to GitHub or third-party storage.
Credentials, don't use: leaked AWS / Stripe / DB creds get one validation call (sts get-caller-identity, banner grab) — never spend, send, or connect.
Redact PII in reports: phone numbers, emails, usernames, tokens, cookies — keep first-2 + last-2, optionally a sha256 fingerprint to prove you had it.
OOB infrastructure: don't hit public DNSLog services. Use the vendor-provided SSRF testing platform if they offer one, or self-host interactsh / your own DNSLog.
No capture, no finding: every claim needs an HTTP capture, screenshot, or video. Don't submit "this should work."

The per-playbook rules are tighter (DoS is the most sensitive, file-upload bans webshells, file-read takes one sample). See the last section of each playbook.

License

MIT. Data sources are public; this project's contribution is the organization, translation to a black-box perspective, and the merge.

6.1 KiB Raw Blame History Unescape Escape