Claude Bug Bounty

Find security vulnerabilities, get paid — with AI doing the heavy lifting

Your AI hunting partner that remembers past targets, spots vulnerabilities, and writes reports for you.

_{by shuvonsec}

What Is This?  ·  Quick Start  ·  Commands  ·  What's New  ·  Install  ·  FAQ  ·  Terms

_{Burp MCP  ·  Caido MCP  ·  HackerOne MCP  ·  Auth-aware hunting  ·  Autonomous mode}

What Is This?

Bug bounty hunting is when companies pay you real money to find security vulnerabilities in their websites and apps before bad actors do. Platforms like HackerOne and Bugcrowd connect hunters with companies. Payouts range from $100 to $1,000,000+ depending on severity.

This tool is a plugin for Claude Code (Anthropic's AI coding assistant) that turns it into a professional bug bounty hunting partner. Instead of juggling 15 different tools and writing reports from scratch, you just type a command and the AI handles the rest.

In plain terms:

• You give it a target website
• It automatically scans the site, finds vulnerabilities, validates they're real, and writes a professional report
• It remembers what you found on past targets and applies that knowledge to new ones
• You can even put it on autopilot and let it hunt on its own while you sleep

Who is it for?

• Security researchers who want to move faster
• Bug bounty hunters who are tired of the manual grind
• People learning security who want AI guidance at every step

Before vs. After

Most hunters waste hours on things that shouldn't take that long. Here's the shift:

Quick Start

Prerequisite: You need Claude Code installed and a Claude Pro or Max plan (or an Anthropic API key with credit). Claude Code itself is free to install, but the underlying model usage requires a paid plan or API billing — the free Claude.ai web account does not include Claude Code access.

Way 1 — Just ask Claude to install it (recommended)

Open your terminal, run claude, and paste this prompt. Claude does the rest:

Help me install the Claude Bug Bounty toolkit from
https://github.com/shuvonsec/claude-bug-bounty into ~/tools/.

Steps:
1. git clone the repo into ~/tools/claude-bug-bounty
2. cd into it and run ./install_tools.sh   (scanners: subfinder, httpx, nuclei, gau, katana, ffuf, dnsx, nmap)
3. then run ./install.sh                   (registers the AI skills + 23 slash commands into ~/.claude/)
4. verify the install by listing the new commands (/recon, /hunt, /validate, /report, /autopilot)
   and confirming the bug-bounty + web2-recon + report-writing skills are available
5. if anything fails (Go missing, Python pkg, brew formula), tell me exactly which tool
   broke and the one-line fix — don't silently skip it

I'm on macOS / Linux. Use brew or apt as needed. Don't sudo unless I approve.

That's it. Claude reads the repo, runs the installers, surfaces any failures, and tells you when you're ready to hunt.

Way 2 — Manual install (if you prefer to drive)

git clone https://github.com/shuvonsec/claude-bug-bounty.git
cd claude-bug-bounty
chmod +x install_tools.sh && ./install_tools.sh   # installs scanning tools (subfinder, httpx, nuclei...)
chmod +x install.sh && ./install.sh               # installs AI skills + commands into Claude Code

Start hunting — paste-ready prompts

Open claude in your terminal and paste any of these. They work as natural-language prompts (no slash commands needed) and let Claude pick the right tools, agents, and gates for the job.

Recon — map the target

Run recon on target.com — subdomain enum, live host probing, URL crawling,
tech fingerprinting, and a focused nuclei sweep. Save everything under
recon/target.com/. Skip any external tool that isn't installed and tell me
what to install to close the gap.

Hunt — test for bugs

Hunt target.com using the recon data already in recon/target.com/.
Prioritize the high-value classes first: IDOR/BOLA, auth bypass, SSRF,
mass assignment, OAuth/OIDC, business logic. Skip anything informational.
Pause before any active POST/PUT/DELETE so I can approve.

Validate — pre-report gate

I think I found a [VULN_CLASS] on [ENDPOINT/PARAM]. Walk it through the
7-Question Gate and the 4 validation gates from skills/triage-validation/.
If it can't clearly pass all 7, kill it — I'd rather drop a maybe than
tank my N/A ratio on the platform.

Report — submission-ready writeup

Write a submission-ready report for the finding I just validated.
Platform: [HackerOne / Bugcrowd / Intigriti / Immunefi].
Use the impact-first template: title formula, one-line impact statement,
exact repro steps, request/response PoC, CVSS 3.1 vector + score,
suggested fix. Human tone — no AI-slop hedging like "could potentially".

Autopilot — full loop, hands off

Run autopilot on target.com in --normal mode. Take it end-to-end:
scope check → recon → attack-surface ranking → hunt → validate → report.
Pause at each gate so I can sanity-check before you spend more credits.
Log everything to hunt-memory so I can resume later.

Pickup — resume a previous hunt

Pick up where I left off on target.com. Read hunt-memory/, show me
untested endpoints from last session, and start with the highest-ranked
ones I haven't touched. Don't re-run recon unless it's older than 7 days.

Smart contract audit (web3)

Audit the Solidity contract at [path/to/Contract.sol]. Walk all 10 DeFi
bug classes from skills/web3-audit: reentrancy, oracle manipulation,
access control, accounting desync, flash loan, signature replay, ERC4626
share inflation, off-by-one, incomplete code paths, proxy/upgrade.
For any finding, drop a Foundry PoC using the template in that skill.

Meme-coin / token rug-pull scan

Scan this token for rug-pull signals: [CHAIN] [contract-address].
Check mint authority, freeze authority, LP lock status, transfer-tax
manipulation, honeypot patterns, bonding-curve exploits, and
proxy/admin keys. Tell me straight: ape, watch, or avoid.

Scope check — before you touch anything

Is [asset.example.com] in scope for the [program-name] bug bounty
program on [HackerOne/Bugcrowd/Intigriti/Immunefi]? Pull the live
policy, check in-scope / out-of-scope lists, flag any caveats
(no automated scanners, no DoS, prod-only, etc.) before I send a
single request.

Intel — what's already known about this target

Pull intel on target.com: relevant CVEs in their tech stack, disclosed
HackerOne / Bugcrowd reports against them or sibling products, recent
patches in their public repos, and any GitHub org I should be watching
for commit leaks.

Tip: Each prompt is also wired to a slash command (/recon, /hunt, /validate, /report, /autopilot, /pickup, /web3-audit, /token-scan, /scope, /intel). Use whichever feels faster — prompts give you control, slash commands are muscle memory.

Don't use Claude Code? Run the Python tools directly:

python3 tools/hunt.py --target target.com
./tools/recon_engine.sh target.com

How It Works

A team of specialists, each doing one job. Claude orchestrates; memory persists across sessions.

flowchart TD
    You([You]) --> Claude[Claude Code]
    Burp[Burp MCP<br/>browser traffic] -.-> Claude
    H1[HackerOne MCP<br/>program intel] -.-> Claude
    Claude --> Recon[Recon<br/>subdomains · URLs · APIs]
    Claude --> Hunt[Hunt<br/>test &amp; validate]
    Claude --> Report[Report<br/>H1 · Bugcrowd · Immunefi]
    Recon --> Memory[(Hunt Memory<br/>persistent across sessions)]
    Hunt --> Memory
    Report --> Memory
    Memory -.feeds back.-> Claude

Run the whole loop, or any step on its own.

Commands

The Core 4 (start here)

Power Commands

Recon Toolkit (v4.3)

Thin wrappers over external tools. Each one is gated on tool presence — missing tools are skipped, not errors.

AI Agents

8 specialized agents, each built for one job:

What's New

v4.3.0 — Auth Sessions + Recon Arsenal (May 2026)

• Auth-aware hunting. Set a session once (--cookie, --bearer, env vars, or .private/target.json) and every downstream tool that takes auth — httpx, katana, ffuf, nuclei, dalfox, the SQLi/SSTI/upload PoC probes — carries it. Most paying bugs (IDOR, BOLA, mass assignment, SSRF behind a login) only exist after login; the default pipeline used to miss them. See docs/auth-sessions.md.
• 8 new commands. /scope-aggregate, /secrets-hunt, /takeover, /cloud-recon, /param-discover, /bypass-403, /scan-cves, /arsenal — all under the Recon Toolkit table below.
• External tool registry. tools/external_arsenal.sh is the single source of truth for ~50 external tools (install hints, upstream URLs, _have <tool> helper). Replaces scattered command -v checks across the shell scripts.
• Recon pipeline. Optional nuclei phase in recon_engine.sh (off by default; gated on tool presence).
• Methodology cheatsheet. skills/security-arsenal/METHODOLOGY_CHEATSHEET.md distills per-vuln quick-check tables from HowToHunt + HolyTips + AllAboutBugBounty + KingOfBugBountyTips into one reference.

v4.2.0 — Memory Rotation (Apr 2026)

• Auto-rotation for hunt memory — audit.jsonl, patterns.jsonl, and journal.jsonl no longer grow forever. Files rotate at 10 MB and keep 3 backups, fully transparent to writers (safe under fcntl.LOCK_EX for concurrent processes).
• /memory-gc — new slash command to inspect, force-rotate, or purge backup files in the hunt-memory tree.
• 22 new tests covering rotation primitives, multi-process concurrent writes, and disk-full OSError propagation.

v4.1.0 — Auto-Memory + README (Apr 2026)

• Auto-memory at session end — the AI now automatically logs what it tested and found after every hunt session. Memory used to stay empty until you manually ran /remember. Now the flywheel starts on day 1.
• README badge and stats updated, install_tools.sh added to Quick Start (was missing)
• hunt-memory/ added to .gitignore (contains full URL history, shouldn't be committed)

v4.0.0 — Meme Coin Security Module (Apr 2026)

• /token-scan <contract> — automated rug pull scanner for EVM and Solana tokens
• skills/meme-coin-audit/ — 8 token bug classes: mint authority, freeze authority, LP locks, honeypot detection, bonding curve exploits, Solana SPL checks
• New agent: token-auditor

v3.1.1 — CI/CD Scanner (Mar 2026)

• GitHub Actions security scanning built into the recon pipeline
• Auto-detects GitHub orgs from recon data and scans their workflow files
• 52 rules, 81.6% GHSA coverage — catches expression injection, secret leaks, supply chain attacks

What It Can Find

Installation

What You Need First

# macOS
brew install go python3 node jq

# Linux (Ubuntu/Debian)
sudo apt install golang python3 nodejs jq

You also need Claude Code installed and a Claude Pro or Max plan (or an Anthropic API key with credit). The free Claude.ai web account does not include Claude Code access — that's the model billing, not the CLI.

Install — two ways

A. Let Claude install it for you. Open claude in your terminal and paste:

Help me install the Claude Bug Bounty toolkit from
https://github.com/shuvonsec/claude-bug-bounty into ~/tools/. Clone the repo,
run ./install_tools.sh, then ./install.sh. Verify the /recon /hunt /validate
/report commands and the bug-bounty skill are registered. If anything breaks
(Go, Python, brew/apt), tell me which tool and the exact one-line fix.

B. Drive it yourself:

git clone https://github.com/shuvonsec/claude-bug-bounty.git
cd claude-bug-bounty
chmod +x install_tools.sh && ./install_tools.sh   # scanning tools (subfinder, httpx, nuclei, etc.)
chmod +x install.sh && ./install.sh               # AI skills + commands into Claude Code

API Keys

export CHAOS_API_KEY="your-key-here"
echo 'export CHAOS_API_KEY="your-key-here"' >> ~/.zshrc

The Rules (Always Active)

These apply every session, no exceptions:

 1. READ FULL SCOPE FIRST   — only test what the program says you can
 2. ONLY REAL BUGS          — "Can an attacker do this RIGHT NOW?" if no, stop
 3. KILL WEAK FINDINGS FAST — 30-second check saves hours of wasted reporting
 4. NEVER GO OUT OF SCOPE   — one wrong request can get you banned
 5. 5-MINUTE RULE           — no progress after 5 min? move to the next target
 6. VALIDATE BEFORE REPORT  — run /validate before you spend 30 min writing
 7. IMPACT FIRST            — start with the bugs that have the worst consequences

Related Projects

Contributing

PRs welcome. Best contributions:

• New vulnerability scanners or detection modules
• Payload additions to skills/security-arsenal/SKILL.md
• Real-world methodology improvements (with evidence from paid reports)
• Support for more platforms (YesWeHack, Synack, HackenProof)

git checkout -b feature/your-contribution
git commit -m "Add: short description"
git push origin feature/your-contribution