bountyforge

All-round bug bounty skill for Claude Code — parallelized agents for smart contract audits (EVM, Move, Solana, TRON), web/API security, local tooling orchestration, and submission-ready reports for HackerOne, Bugcrowd, Intigriti & Immunefi.

AI Pentesting Tool: Run isolated, cloud-hosted pentesting sandboxes at bountyforge.xyz — your AI key, your Firecracker microVM, your report. Pipeline: recon → hunt → triage → H1-ready report. AI slop gets you rate-limited; BountyForge gets you paid.

New in v3.1.0: Al-Mizaan v3 deep validation gates, SIS-MD passive intelligence integration, agent isolation checks. See CHANGELOG.md.

What It Does

Bounty Forge spins up 8 specialized security agents in parallel, each attacking a different surface of your target. Findings are deduplicated, gate-evaluated, CVSS-scored, and formatted into a submission-ready report — in minutes.

Supported targets: web/API, smart contracts, infrastructure, supply chain, internal tooling, binary analysis

Local tooling: When Claude Code execution is enabled, BountyForge can orchestrate local CLI tools like nmap, ffuf, amass, sqlmap, gobuster, curl, httpx, wfuzz, zap, burpsuite, and other installed scanners/fuzzers.

Payload coverage: Designed to explore unlimited payload variants for SQL injection, CSV injection, open redirect, XSS, SSRF, command injection, template injection, path traversal, deserialization, prototype pollution, auth bypass, business logic abuse, IDOR, CSRF, response splitting, and more.

Reference setup files: references/setup.md and references/local-tooling.md contain the actual Deepseek CLI, local tooling, and vulnerability environment instructions the skill uses.

Report formats: HackerOne · Bugcrowd · Intigriti · Immunefi · Generic

AI Pentesting Tool — bountyforge.xyz

Take BountyForge hunting without spinning up your own environment. bountyforge.xyz hosts the same engine in isolated pentesting sandboxes:

• Firecracker microVM per session — your AI key, your box, your report
• Full recon → hunt → triage → H1-ready report pipeline in the browser
• No rate limits, no AI slop — clean, isolated execution
• Hosted skill releases managed from this repo

Choose your platform: install the open-source skill here, or hunt in the cloud at bountyforge.xyz.

Installation

Claude Code (terminal)

git clone https://github.com/Gabson0x/bountyforge.git ~/.claude/skills/bountyforge

Start a fresh Claude Code session — skills load at startup.

Claude.ai (web/app)

1. Go to Customize → Skills
2. Make sure Code execution is enabled in Settings → Capabilities
3. Upload the .skill file from Releases

Optional: Bug Bounty Intelligence MCP (smart contract scanning)

For automated Solidity scanning with Al-Mizaan v3 7-gate analysis, add the companion MCP server:

claude mcp add bug-bounty-intelligence -- npx -y bug-bounty-intelligence-mcp@latest

Or in ~/.claude/settings.json:

{
  "mcpServers": {
    "bug-bounty-intelligence": {
      "command": "npx",
      "args": ["-y", "bug-bounty-intelligence-mcp@latest"]
    }
  }
}

Tools available:

BountyForge auto-detects the MCP and uses list_vulnerability_patterns (free) for pre-hunt bug-class prioritization. See references/bug-bounty-intelligence-mcp.md for full integration guide.

Enabling Claude Code Execution (local execution)

To allow BountyForge to run local tools and subagents, enable Claude Code (local execution) in your Claude environment and grant the skill permission to execute shell commands.

macOS / Linux (Claude Code client):

1. Start Claude Code with code-execution enabled (follow your Claude Code client docs).
2. Ensure the shell that launches Claude Code has the tools you want on PATH (e.g., nmap, ffuf, sqlmap).
3. If using project-specific env vars, source your project file before starting Claude Code:

source .env            # or project.env
claude start           # or the command your Claude Code client uses

Windows (Claude.app / PowerShell):

1. Open PowerShell as the user that runs Claude.
2. Set any project env vars or tokens (example shown in references/setup.md).
3. Launch Claude with code execution enabled from the same session so it inherits the environment.

Notes:

• If you are using the web/app variant, go to Settings → Capabilities and toggle Code execution on. Some deployments require you to enable a "subagent" or "local tooling" checkbox; consult your Claude distribution docs.
• Always start Claude from the shell that has your project environment loaded so deepseek export, nmap, and other tools are available to the skill.
• For privacy and safety, grant execution permission only for trusted skills and projects.

Deepseek Pro Setup (Claude CLI)

BountyForge includes references/setup.md so the skill can use the same Deepseek CLI environment and local tooling configuration during audit runs.

Mac / Linux

In your project shell or project file, export:

export ANTHROPIC_BASE_URL=https://api.deepseek.com/anthropic
export ANTHROPIC_MODEL=deepseek-v4-pro
export ANTHROPIC_DEFAULT_OPUS_MODEL=deepseek-v4-pro
export ANTHROPIC_DEFAULT_SONNET_MODEL=deepseek-v4-pro
export ANTHROPIC_DEFAULT_HAIKU_MODEL=deepseek-v4-flash
export CLAUDE_CODE_SUBAGENT_MODEL=deepseek-v4-flash
export CLAUDE_CODE_EFFORT_LEVEL=max
export ANTHROPIC_AUTH_TOKEN="your-deepseek-pro-token"

Then bind the current repo:

deepseek export --project . --key "$ANTHROPIC_AUTH_TOKEN" --mode pro

Start Claude CLI from the same shell so the environment variables are active.

Windows (PowerShell)

Use these variables in the current session:

$env:ANTHROPIC_BASE_URL = "https://api.deepseek.com/anthropic"
$env:ANTHROPIC_MODEL = "deepseek-v4-pro"
$env:ANTHROPIC_DEFAULT_OPUS_MODEL = "deepseek-v4-pro"
$env:ANTHROPIC_DEFAULT_SONNET_MODEL = "deepseek-v4-pro"
$env:ANTHROPIC_DEFAULT_HAIKU_MODEL = "deepseek-v4-flash"
$env:CLAUDE_CODE_SUBAGENT_MODEL = "deepseek-v4-flash"
$env:CLAUDE_CODE_EFFORT_LEVEL = "max"
$env:ANTHROPIC_AUTH_TOKEN = "your-deepseek-pro-token"

deepseek export --project . --key $env:ANTHROPIC_AUTH_TOKEN --mode pro

For persistence, add the same variables to your PowerShell profile.

Note: this setup is intended for temporary use inside a project or shell session so the Deepseek CLI export command can be applied without modifying the core skill files.

Usage

How to use Bounty Forge

1. Choose a precise scope. For bug bounty work, point the skill at the exact contract file, API endpoint list, or web target you are testing.
2. Enable code execution. When Claude Code can run locally, Bounty Forge can orchestrate local tools like nmap, ffuf, amass, sqlmap, gobuster, curl, httpx, and zap.
3. Run an audit command. Use the examples below and include --file-output or --cvss when you want a formatted deliverable.
4. Review findings. The skill will classify and gate each finding, but always validate exploitability before submitting.
5. Generate a report. Use the built-in report mode to turn confirmed findings into platform-ready output.

Audit a contract or repo

audit contracts/

run bountyforge on src/usdc.move --platform immunefi --cvss --file-output

/bountyforge

check this contract for vulns --platform h1 --cvss

Web / API target

/bountyforge on https://api.target.com

find vulns in this API — [paste endpoints / Swagger / JS bundle]

Generate a report from findings

write a HackerOne report for this finding: [paste notes]

generate immunefi report --cvss: [describe the vuln]

Flags

How It Works

Discover files / scope
        ↓
Build agent bundles (source + agent instructions)
        ↓
Spawn 8 agents in parallel
        ↓
Deduplicate findings by (Target | location | bug-class)
        ↓
Gate evaluation: Refutation → Reachability → Trigger → Impact
        ↓
CVSS 3.1 scoring
        ↓
Submission-ready report

Every finding passes four gates before it's confirmed:

1. Refutation — can the attack be concretely blocked by an existing guard?
2. Reachability — can the vulnerable state exist in a live deployment?
3. Trigger — can an unprivileged actor execute it profitably?
4. Impact — is there material harm to an identifiable victim?

Fail any gate → rejected or demoted to a lead for manual review.

Tips

• Target hot contracts. Point BountyForge at the 2–5 files you're actively reviewing rather than an entire repo. Smaller scope = denser context per agent = higher-signal findings.
• Run more than once. LLM output is non-deterministic — each run can surface different vulnerabilities. Two or three passes often catch what a single pass misses.
• Chain findings. BountyForge automatically detects composite chains (e.g., auth bypass → privilege escalation) and reports them as a combined severity.
• Use --file-output when submitting. It saves a clean markdown report you can paste directly into the platform.

Updating

cd ~/.claude/skills/bountyforge
git pull

BountyForge checks for updates automatically on each run and will warn you if a newer version is available.

Structure

bountyforge/
├── SKILL.md                          # Main orchestrator
├── VERSION                           # Current version
├── CHANGELOG.md                      # Release notes
└── references/
    ├── judging.md                    # 4-gate evaluation rules
    ├── supervisor.md                 # Detailed triage supervisor system
    ├── knowledge.md                  # Disclosed-report knowledge base
    ├── report-formatting.md          # Platform report templates
    ├── cvss-guide.md                 # CVSS 3.1 scoring guide
    ├── setup.md                      # Deepseek CLI environment guidance
    ├── local-tooling.md              # Local tooling and vuln coverage reference
    ├── al-mizaan-gates.md            # Al-Mizaan v3 7-gate deep validation (from Bug Bounty Intelligence MCP)
    ├── sis-intelligence.md           # SIS-MD passive security intelligence integration
    ├── isolation.md                  # Agent isolation rules and boundary enforcement
    ├── cwe-knowledge-base.md          # 1,047 CWEs across 16 agent domains with detection patterns
    ├── attack-vectors/
    │   ├── web-api-vectors.md
    │   ├── smart-contract-vectors.md
    │   ├── business-logic-vectors.md
    │   └── ...
    └── hacking-agents/
        ├── shared-rules.md
        ├── web-api-agent.md
        ├── smart-contract-agent.md
        ├── access-control-agent.md
        ├── business-logic-agent.md
        ├── crypto-math-agent.md
        ├── race-condition-agent.md
        ├── economic-security-agent.md
        └── recon-agent.md

Collaboration & Integrated Projects

BountyForge v3.1.0 integrates methodologies from:

Key lessons integrated:

• From Bug Bounty Intelligence: 100% of Slither's "High" findings on a mature protocol were false positives — 89% were out-of-scope lib/ noise. Scope-aware, context-aware analysis eliminates this class of error entirely.
• From SIS-MD: Passive analysis before active hunting catches metadata leaks, secret sprawl, and technology misconfigurations that active scanners miss. Masked reporting prevents the report itself from becoming a leak vector.

Disclaimer

BountyForge is intended for authorized security research and bug bounty programs only. Only use it against targets you have explicit permission to test. The author is not responsible for misuse.

Built by @Gabson0x · Synq Studio