A Claude Code skill bundle for bug hunting and external red-team work - 82 skills, 15 slash commands, 681 disclosed-report patterns curated across 24 core vulnerability classes, plus enterprise identity + infrastructure attack matrices.
Python ai-securityanthropicapplication-securitybug-bountybugbountybugcrowdclaudeclaude-codeclaude-skillsethical-hackinghackeroneoffensive-securitypentestingred-teamsecurity-toolsweb-security
Phase 3.6: Positioning fix — slash commands PRIMARY, cbh CLI SECONDARY
Sachin Sharmadb0413e555b1
diff --git a/README.md b/README.mdindex 3647279..83df579 100644--- a/README.md+++ b/README.md@@ -25,6 +25,28 @@ All triggered automatically by topic — describe what you're testing in plain E---+## Two interfaces — pick what fits your engagement++The bundle exposes the same content through two interfaces. **Slash commands are the primary interface**; the `cbh` CLI is a secondary terminal-native runner. Both consume the same `skills/` content; they differ in execution model.++| | Slash commands (PRIMARY) | `cbh` CLI (SECONDARY) |+|---|---|---|+| Runs in | A Claude Code conversation | Any terminal with Python 3.9+ |+| Execution | LLM-driven — reads full SKILL.md, applies judgment, can chain skills, can converse | Deterministic — Python stdlib, regex match, real `subfinder`/`dig`/`curl` calls |+| Output | Conversational, contextual, varies per run | Files + structured stdout, identical across runs |+| Best for | Hunting, chain construction, applying discipline rules with nuance, talking through findings | CI/CD, scripted automation, bulk recon, deterministic verification, non-Claude environments |+| Examples | `/recon target.com` `/hunt target.com` `/triage` `/report` `/validate` `/chain` `/autopilot` `/scope` | `cbh recon target.com` (real network I/O) · `cbh triage finding.md` (deterministic 7Q grep) · `cbh report finding.md --platform bugcrowd` |++**Choose by use case:**++- **Exploring a new target?** Use Claude Code with slash commands. The LLM applies judgment that the deterministic CLI can't.+- **Running scheduled recon? Verifying labs? CI gate?** Use `cbh`. It's reproducible and scriptable.+- **You don't have Claude Code installed but want to read the skills/Pattern Libraries?** Use `cbh` plus `cat skills/<name>/SKILL.md`. The content stands on its own.++See [`docs/cbh-cli.md`](docs/cbh-cli.md) for the CLI reference. See the slash command list under **Slash Commands** later in this file for the conversational interface.++---+## Structure```diff --git a/docs/cbh-cli.md b/docs/cbh-cli.mdindex d77cd80..adb8ad6 100644--- a/docs/cbh-cli.md+++ b/docs/cbh-cli.md@@ -1,10 +1,28 @@# `cbh` — claude-bughunter CLI-> Native runner that bridges the skill content into a real engagement loop.-> Four subcommands compose intake → reconnaissance → triage → submission.+> **Secondary interface — slash commands are primary.** Inside a Claude Code conversation, use the slash commands (`/recon`, `/hunt`, `/triage`, `/report`, `/validate`, `/chain`, `/autopilot`, `/scope`, etc.) — they leverage the full skill content and the LLM's judgment.+>+> `cbh` is the **terminal-native deterministic runner** — use it when you're outside Claude Code, automating in CI/CD, running scheduled recon, or verifying labs reproducibly. Same skills, different execution model.++## When to use `cbh` vs slash commands++| Use case | Use this |+|---|---|+| Hunting a new target conversationally, applying judgment | **Slash commands** in Claude Code (`/hunt`, `/triage`, etc.) |+| Building a chain across multiple primitives | **Slash commands** — LLM keeps state across the conversation |+| Scheduled / CI / scripted runs | **`cbh`** — deterministic exit codes, identical output across runs |+| Bulk passive recon (hundreds of subdomains) | **`cbh recon`** — real `subfinder`/`dig`/`curl`, no LLM in the loop |+| Verifying labs / reproducing claims | **`cbh`** — every Phase 2 doc's curls work via `cbh` too |+| Reading skills without Claude Code installed | **`cbh`** + browsing `skills/` and `docs/disclosed-reports/` |+| Triage gate at PR time / pre-submit linting | **`cbh triage`** — deterministic keyword-match against the 7-Question Gate |++The two interfaces consume the same content (`skills/` + `docs/disclosed-reports/`). They produce different outputs because they execute differently. Pick by context, not by preference.++## Operating modes for the CLI+> Stdlib + optional `subfinder` for richer recon. No build step.>-> **Two operating modes** — pick what fits your setup:+> **Two HTTP-routing modes** within the CLI — pick what fits your setup:> 1. **Curl-only (default)** — stdlib HTTP, no Burp dependency. Works on any laptop with Python 3.9+.> 2. **Burp Suite integration** — `--burp` flag routes everything through Burp's proxy (default `127.0.0.1:8080`). Requests + responses land in Proxy → HTTP history; you can send any of them to Repeater/Intruder/Scanner/Collaborator. Pairs with the **Burp MCP server** (port 9876) for Claude-Code-conversational hunting.diff --git a/scripts/cbh.py b/scripts/cbh.pyindex e1f59ab..b7fee13 100755--- a/scripts/cbh.py+++ b/scripts/cbh.py@@ -295,7 +295,7 @@ def cmd_recon(args: argparse.Namespace) -> int:say(f" HTTP-live: {len(live)}")say(f" Output: {out_dir}")say()- say(f" Next: {color(f'cbh classify <url>', 'bold')} to pattern-match attack candidates")+ say(f" Next: {color('cbh classify <url>', 'bold')} for fast pattern-match, or {color('/hunt <target>', 'bold')} in Claude Code for full LLM-driven hunting")return 0@@ -449,6 +449,10 @@ def cmd_classify(args: argparse.Namespace) -> int:say(f" 2. Pick the highest-confidence attack from the Pattern Library")say(f" 3. Apply OOB-Or-It-Didn't-Happen Gate before claiming success")say(f" 4. {color('cbh triage <finding.md>', 'bold')} once you have a candidate finding")+ say()+ say(color("Need richer context? In a Claude Code conversation:", "dim"))+ say(color(f" /hunt <url> — full hunt-dispatch routing with LLM judgment", "dim"))+ say(color(f" /chain — build A→B→C exploit chains", "dim"))return 0@@ -682,14 +686,22 @@ def cmd_report(args: argparse.Namespace) -> int:def main() -> int:parser = argparse.ArgumentParser(prog="cbh",- description="claude-bughunter CLI — orchestrate the engagement loop",+ description=("claude-bughunter CLI — terminal-native deterministic runner.\n"+ "SECONDARY interface; slash commands (/hunt, /recon, /triage, /report) "+ "in Claude Code are PRIMARY. Use cbh for CI/CD, scripted runs, "+ "deterministic verification, or when not in a Claude Code session."),formatter_class=argparse.RawDescriptionHelpFormatter,epilog=textwrap.dedent("""\Examples:cbh recon hackerone.com+ cbh recon target.com --burp # route via Burp proxycbh classify "https://api.target.com/v1/users/42?next=https://evil.com"cbh triage findings/idor-2026-05-15.mdcbh report findings/idor-2026-05-15.md --platform bugcrowd --out draft.md++ For LLM-driven hunting with full skill context, use the slash commands+ inside Claude Code: /hunt /recon /triage /report /validate /chain /autopilot+ See docs/cbh-cli.md for the "when to use which" matrix."""),)sub = parser.add_subparsers(dest="command", required=True)