VulnHunter Security Scanner

Skill by ara.so — Security Skills collection.

VulnHunter is an agentic AI security tool that applies proactive, attacker-first analysis directly to source code. Unlike traditional SAST scanners that flag patterns and generate false positives, VulnHunter reasons like an adversary: it identifies which defects are actually exploitable, maps attack paths, and proposes evidence-backed fixes.

Core Concepts

Attacker-First Forward Analysis: Starts at attacker-accessible entry points (APIs, file uploads, network messages) and reasons forward to evaluate whether an attacker can truly break through, rather than working backward from dangerous sinks.

Falsification Engine: After finding a potential vulnerability, VulnHunter runs a structured workflow to disprove its own argument, searching for flawed assumptions, logic gaps, or security controls that would block the attack.

Three-Skill Loop:

/vulnhunt - Hunt for vulnerabilities
/vulnhunter-fix - Fix verified issues with test-driven remediation
/vulnhunt-fix-verify - Independently verify fixes

Prerequisites

Claude Code CLI authenticated with access to Claude Opus (required)
Python 3.12+ (for agent runtime and harness tooling)
Git and GitHub CLI (gh) authenticated (for fix workflow)
Anthropic Cyber Verification Program enrollment recommended to avoid cyber safeguard blocks

Installation

# Clone the repository
git clone https://github.com/capitalone/vulnhunter.git
cd vulnhunter

# Install skills to ~/.claude/skills/
./install.sh

# Optional: Uninstall
# ./uninstall.sh

Note: install.sh copies files (not symlinks) because symlinks can break subagent functionality. Re-run after pulling updates.

Core Usage

1. Running the Scanner (`/vulnhunt`)

# Launch Claude Code with the vulnhunt skill
claude --model opus \
  --add-dir ~/.claude/skills/vulnhunt \
  --add-dir ~/.claude/skills/vulnhunt/phases

# Inside Claude Code session:
/vulnhunt

The scanner will:

Recon Phase: Map entry points and dangerous sinks
Parallel Hunt: Trace paths from entries to sinks
Adversarial Disprove: Apply falsification to eliminate false positives
Capability Filter: Verify actual exploitability
Output: Emit only verified issues with exploit paths and proposed fixes

2. Running the Fixer (`/vulnhunter-fix`)

First install Python dependencies:

cd vulnhunter-fix
pip install -e ".[dev]"

Launch with the fix skill:

claude --model opus --add-dir ~/.claude/skills/vulnhunter-fix

# Inside Claude Code session:
/vulnhunter-fix

The fixer implements RED-GREEN-REFACTOR workflow:

Write exploit demonstration
Create failing security test (RED)
Implement code fix (GREEN)
Verify exploit blocked without regressions
Cut reviewable PR

3. Running the Verifier (`/vulnhunt-fix-verify`)

The verifier runs read-only validation on fixes. Pre-create the output directory:

mkdir -p /tmp/verify-output

claude --model opus \
  --add-dir ~/.claude/skills/vulnhunt-fix-verify \
  --add-dir ~/.claude/skills/vulnhunt-fix-verify/phases

# Inside Claude Code session:
/vulnhunt-fix-verify \
  repo=/absolute/path/to/repo \
  report=/absolute/path/to/vulnhunt_report.md \
  fixed=VULN-001,VULN-003 \
  out=/tmp/verify-output \
  comments=/absolute/path/to/pr_comments.json \
  additional_repos=/path/to/lib1,/path/to/lib2

Parameters:

repo - Absolute path to repository root
report - Absolute path to VulnHunt findings report
fixed - Comma-separated list of vulnerability IDs to verify
out - Absolute path to output directory (must exist)
comments - (Optional) Path to PR review comments JSON
additional_repos - (Optional) Comma-separated paths to dependency repos

Headless Runtime Agent

For CI/CD or non-interactive pipelines:

cd vulnhunter-agent
pip install -e ".[dev]"

# Configure via config.yaml
cp config.example.yaml config.yaml
# Edit config.yaml with your settings

Example config.yaml:

anthropic_api_key: ${ANTHROPIC_API_KEY}
model: claude-opus-4-20250514
github_token: ${GITHUB_TOKEN}

targets:
  - url: https://github.com/org/repo
    branch: main
    language: python
    create_issues: true

output_dir: ./scan-results
max_concurrent: 3

Run headless scan:

python -m vulnhunter_agent.main --config config.yaml

The agent will:

Clone target repositories
Execute /vulnhunt skill
Parse findings
Create GitHub issues for verified vulnerabilities

Batch Scanning with Harness

For workstation-scale batch operations:

cd harness
pip install -e ".[dev]"

# Configure target list
echo "https://github.com/org/repo1" >> local_harness/batch/REPO_LIST.txt
echo "https://github.com/org/repo2" >> local_harness/batch/REPO_LIST.txt
# Lines starting with # are ignored

# Run batch scan
python -m local_harness.batch.run scan

# Resume interrupted scan
python -m local_harness.batch.run scan --resume

# Check progress
python -m local_harness.batch.run status

# Collect all findings
python -m local_harness.batch.run collect

Results are stored in harness/local_harness/batch/workdir/.

Benchmarking Mode

Evaluate scanner accuracy against known vulnerabilities:

cd harness
pip install -e ".[dev]"

# Run full benchmark
python -m local_harness.benchmark.run

# Benchmark specific repository
python -m local_harness.benchmark.run --repos "OWASP/NodeGoat"

# Regenerate metrics only
python -m local_harness.benchmark.run --tally-only

Creating Ground Truth

Define known vulnerabilities in harness/local_harness/benchmark/ground_truth/<repo>.json:

{
  "repo_name": "my-vulnerable-app",
  "repo_url": "https://github.com/org/my-vulnerable-app",
  "vulnerabilities": [
    {
      "id": "SQL-001",
      "type": "SQL Injection",
      "file": "src/db/user.py",
      "line": 42,
      "severity": "critical",
      "description": "Unsanitized user input in SQL query",
      "exploit_vector": "username parameter in /api/login"
    }
  ]
}

Configure benchmark engines in harness/local_harness/config.py:

BENCHMARK_CONFIG = {
    "scanner_model": "claude-opus-4-20250514",
    "judge_model": "claude-opus-4-20250514",
    "max_workers": 4
}

Configuration

VulnHunt Scanner Configuration

The scanner is prompt-driven. Customize behavior by editing vulnhunt/SKILL.md and phase files in vulnhunt/phases/:

vulnhunt/phases/recon.md - Entry point discovery
vulnhunt/phases/hunt.md - Path tracing logic
vulnhunt/phases/disprove.md - Falsification rules
vulnhunt/phases/capability.md - Exploitability verification

VulnHunter-Fix Configuration

Edit vulnhunter-fix/vulnhunter_fix/config.py:

DEFAULT_CONFIG = {
    "test_framework": "pytest",  # or "unittest", "jest", etc.
    "create_pr": True,
    "pr_base_branch": "main",
    "require_tests": True,
    "max_fix_attempts": 3
}

Or override at runtime by creating .vulnhunter-fix.yaml in project root:

test_framework: pytest
create_pr: true
pr_base_branch: develop
require_tests: true
max_fix_attempts: 5

Common Patterns

Pattern 1: Quick Security Audit

# One-shot vulnerability scan
claude --model opus \
  --add-dir ~/.claude/skills/vulnhunt \
  --add-dir ~/.claude/skills/vulnhunt/phases \
  << 'EOF'
/vulnhunt
EOF

Pattern 2: Full Remediation Workflow

#!/usr/bin/env python3
"""Complete hunt-fix-verify workflow"""
import subprocess
import json
from pathlib import Path

def run_workflow(repo_path: Path):
    # Step 1: Scan
    print("[1/3] Scanning for vulnerabilities...")
    subprocess.run([
        "claude", "--model", "opus",
        "--add-dir", "~/.claude/skills/vulnhunt",
        "--add-dir", "~/.claude/skills/vulnhunt/phases",
        "-c", "/vulnhunt"
    ], cwd=repo_path)
    
    # Step 2: Fix
    print("[2/3] Applying fixes...")
    subprocess.run([
        "claude", "--model", "opus",
        "--add-dir", "~/.claude/skills/vulnhunter-fix",
        "-c", "/vulnhunter-fix"
    ], cwd=repo_path)
    
    # Step 3: Verify
    print("[3/3] Verifying fixes...")
    out_dir = Path("/tmp/verify-out")
    out_dir.mkdir(exist_ok=True)
    
    subprocess.run([
        "claude", "--model", "opus",
        "--add-dir", "~/.claude/skills/vulnhunt-fix-verify",
        "--add-dir", "~/.claude/skills/vulnhunt-fix-verify/phases",
        "-c", f"/vulnhunt-fix-verify repo={repo_path} report={repo_path}/vulnhunt_report.md out={out_dir}"
    ])

if __name__ == "__main__":
    run_workflow(Path.cwd())

Pattern 3: CI/CD Integration

# .github/workflows/vulnhunt.yml
name: VulnHunter Security Scan

on:
  schedule:
    - cron: '0 2 * * 1'  # Weekly Monday 2 AM
  workflow_dispatch:

jobs:
  security-scan:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v4
      
      - name: Setup Python
        uses: actions/setup-python@v5
        with:
          python-version: '3.12'
      
      - name: Install VulnHunter Agent
        run: |
          git clone https://github.com/capitalone/vulnhunter.git
          cd vulnhunter/vulnhunter-agent
          pip install -e ".[dev]"
      
      - name: Run Security Scan
        env:
          ANTHROPIC_API_KEY: ${{ secrets.ANTHROPIC_API_KEY }}
          GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
        run: |
          cd vulnhunter/vulnhunter-agent
          cat > config.yaml << EOF
          anthropic_api_key: ${ANTHROPIC_API_KEY}
          github_token: ${GITHUB_TOKEN}
          model: claude-opus-4-20250514
          targets:
            - url: ${{ github.server_url }}/${{ github.repository }}
              branch: ${{ github.ref_name }}
              create_issues: true
          output_dir: ./results
          EOF
          python -m vulnhunter_agent.main --config config.yaml
      
      - name: Upload Results
        uses: actions/upload-artifact@v4
        with:
          name: vulnhunt-results
          path: vulnhunter/vulnhunter-agent/results/

Pattern 4: Custom Falsification Rules

Extend the falsification engine by adding constraints to vulnhunt/phases/disprove.md:

## Custom Falsification Rules

### Framework-Specific Protections

**Django ORM**: If the code uses Django ORM methods like `.filter()`, 
`.get()`, `.exclude()` with keyword arguments (not raw SQL), the ORM 
provides automatic parameterization. Flag this as NOT exploitable.

**Express.js Helmet**: If `helmet()` middleware is configured and the 
vulnerability is XSS-related, verify CSP headers block inline script 
execution.

### Language-Specific Checks

**Python Type Hints**: If function signatures use strict type hints 
with runtime validation (e.g., Pydantic), trace whether malicious input 
can bypass type coercion.

**Rust Ownership**: For memory safety issues, verify whether Rust's 
borrow checker already prevents the exploit at compile time.

Testing

Each component has isolated test suites:

# Test harness
cd harness
pip install -e ".[dev]"
python -m pytest tests/ --cov=local_harness

# Test fixer
cd vulnhunter-fix
pip install -e ".[dev]"
python -m pytest -q

# Test agent
cd vulnhunter-agent
pip install -e ".[dev]"
python -m pytest -q

Troubleshooting

Issue: Cyber Safeguard Blocks

Symptom: Requests blocked with cyber abuse warnings

Solution: Enroll in Anthropic Cyber Verification Program

# Verify enrollment status
curl https://api.anthropic.com/v1/cyber/status \
  -H "x-api-key: $ANTHROPIC_API_KEY"

Issue: High False Positive Rate

Symptom: Many reported vulnerabilities are not exploitable

Solution: Strengthen falsification phase:

Review vulnhunt/phases/disprove.md
Add domain-specific sanitization patterns
Increase falsification iterations in SKILL.md:

## Falsification Configuration

Run **3 independent falsification passes** (up from default 1):
- Pass 1: Check input validation
- Pass 2: Verify sanitization/encoding
- Pass 3: Confirm no bypass via edge cases

Issue: Scanner Misses Known Vulnerabilities

Symptom: Benchmark shows low recall

Solution: Expand reconnaissance phase:

Edit vulnhunt/phases/recon.md to include more entry point types
Add framework-specific entry points:

## Additional Entry Points

**FastAPI**: `@app.post()`, `@app.get()` decorated async functions
**Spring Boot**: `@RestController`, `@RequestMapping` methods
**GraphQL**: `resolve_*` functions in schema definitions

Issue: Verifier Rejects Valid Fixes

Symptom: /vulnhunt-fix-verify marks fixes as unsuccessful

Solution: Check verification logs for specific failure reason:

cat /path/to/verify-output/VULN-001-verification.log

Common causes:

Test not exercising exact exploit path
Fix incomplete (e.g., sanitizes one parameter but not all)
Regression introduced (unrelated functionality broken)

Issue: Headless Agent Crashes

Symptom: vulnhunter-agent exits with errors

Solution: Enable debug logging:

# config.yaml
log_level: DEBUG
log_file: ./agent-debug.log

Check common issues:

Invalid GitHub token: gh auth status
API rate limits: curl -H "x-api-key: $ANTHROPIC_API_KEY" https://api.anthropic.com/v1/usage
Disk space: df -h

Issue: Skill Not Found

Symptom: /vulnhunt command not recognized

Solution: Verify installation:

ls -la ~/.claude/skills/vulnhunt
ls -la ~/.claude/skills/vulnhunt/phases

# Reinstall if missing
cd /path/to/vulnhunter
./install.sh

Environment Variables

# Required for Anthropic API access
export ANTHROPIC_API_KEY="your-api-key"

# Required for GitHub integration
export GITHUB_TOKEN="your-github-token"

# Optional: Custom skill directory
export CLAUDE_SKILLS_DIR="$HOME/custom-skills"

# Optional: Control batch concurrency
export VULNHUNT_MAX_WORKERS=4

# Optional: Verification timeout (seconds)
export VULNHUNT_VERIFY_TIMEOUT=1800

Additional Resources

Official Docs: VulnHunter GitHub
Claude Code Docs: docs.claude.com/claude-code
Cyber Verification: Anthropic CVP Portal
Contributing: See CONTRIBUTING.md in repository
Security: Report issues per SECURITY.md guidelines

vulnhunter-security-scanner