Skip to content

ByamB4/find-cve-agent

Repository files navigation

find-cve-agent

Open Source CVE Hunting Harness for Claude Code

A Claude Code plugin that systematically finds real CVEs in open source packages through coordinated multi-agent security research.


What It Is

find-cve-agent is a battle-tested harness of 20 skills organized as a 5-agent team. It provides structured workflows for every phase of vulnerability research: target discovery, code review, PoC development, false positive elimination, and responsible disclosure.

Every skill encodes practical knowledge about what gets accepted, what gets rejected, and how to avoid wasting time on false positives.

Philosophy

  • Quality over quantity. One confirmed CVE beats ten false positives.
  • False positive elimination is a first-class concern. Every finding passes a 6-gate verification process before submission.
  • Responsible disclosure only. 90-day coordinated timeline, no production exploitation, PoCs run locally.
  • Learn from mistakes. The plugin encodes patterns from past false positives so you don't repeat them.

Quick Start

# Clone the plugin
git clone https://github.com/find-cve-agent/find-cve-agent.git

# Install into your project
cd your-research-workspace
bash /path/to/find-cve-agent/install.sh

# Start hunting
# In Claude Code:
/hunt <package-name>

Or manually copy the plugin contents into your project's .claude/ directory.

Recommended: Also install blader/humanizer to auto-clean AI patterns from disclosure reports before sending.


Agent Team

Agent Role Key Responsibility
Director Human lead Approves targets, approves PoC plans, final submit/drop
Recon Target discovery Finds promising packages on npm/PyPI/GitHub
Hunter Code review Traces data flows from source to sink
Exploiter PoC builder Writes exploits, chains findings for max impact
Validator FP eliminator 6-gate verification, kills false positives
Registry Bookkeeper Tracks all targets, prevents duplicate work

Available Commands

Command Description
/hunt <package> Full pipeline: registry check -> clone -> review -> PoC -> validate -> report
/recon <category> Find targets in a category (e.g., "csv parsers", "template engines")
/check-nvd <package> Query NVD and OSV.dev for existing CVEs
/fp-check Run the 6-gate false positive elimination on current finding
/report Generate a disclosure report from current finding
/registry [query] Query or update the research registry
/cross-pollinate Find the same vulnerability pattern in similar packages

Available Skills

Reconnaissance

  1. Target Discovery — Find under-audited packages with high attack surface
  2. Download Analysis — Assess real-world usage via registry download counts
  3. Advisory Mining — Find incomplete fixes for existing CVEs
  4. Attack Surface Mapping — Identify entry points and data flows

Hunting

  1. Command Injection Search — Find exec/spawn/system sinks
  2. Path Traversal Search — Find file write operations with user paths
  3. Template Injection Search — Find compile-from-string patterns
  4. Deserialization Search — Find unsafe object reconstruction
  5. SSRF Search — Find user-controlled URL fetching
  6. XXE Search — Find XML parsing without entity limits
  7. SQL Injection Search — Find string concatenation in queries
  8. Auth Bypass Search — Find missing or flawed auth checks
  9. ReDoS Search — Find exponential-backtracking regex patterns
  10. Prototype Pollution Search — Find unsafe merge/clone/assign

Exploitation

  1. PoC Builder — Structured exploit script generation
  2. Chain Escalation — Combine findings for higher impact
  3. CVSS Calculator — Accurate severity scoring

Validation

  1. FP-Check (6 Gates) — Systematic false positive elimination
  2. Version Verification — Confirm vuln exists in latest release

Reporting

  1. Disclosure Report Generator — Format findings for submission channels

Directory Structure

find-cve-agent/
+-- .claude-plugin/
|   +-- plugin.json          # Plugin manifest
+-- CLAUDE.md                # Agent architecture and instructions
+-- README.md                # This file
+-- LICENSE                  # Apache-2.0
+-- install.sh               # Installation script
+-- agents/
|   +-- recon.md             # Target discovery agent
|   +-- hunter.md            # Code review agent
|   +-- exploiter.md         # PoC builder agent
|   +-- validator.md         # FP elimination agent
|   +-- registry.md          # Tracking agent
+-- commands/
|   +-- hunt.md              # /hunt command
|   +-- recon.md             # /recon command
|   +-- check-nvd.md         # /check-nvd command
|   +-- fp-check.md          # /fp-check command
|   +-- report.md            # /report command
|   +-- registry.md          # /registry command
|   +-- cross-pollinate.md   # /cross-pollinate command
+-- hooks/
|   +-- hooks.json           # Hook configuration
|   +-- session-start-context.mjs
|   +-- pretooluse-clone-dedup.mjs
|   +-- pretooluse-finding-selfcheck.mjs
|   +-- posttooluse-version-check.mjs
+-- templates/
    +-- REGISTRY.md          # Empty registry template

Environment Requirements

  • Required: git, gh (GitHub CLI), python3, node, curl
  • Optional: npm, pip3 (for target-specific testing)
  • Platform: macOS or Linux
  • Claude Code: Latest version with plugin support

Contributing

Contributions welcome. Areas of interest:

  • New vulnerability search patterns — Add skills for emerging vuln classes
  • False positive patterns — Document new FP patterns you've encountered
  • Disclosure templates — Improve report formatting for different channels
  • Agent improvements — Better prompts, better workflows

Please open an issue first to discuss significant changes.


License

Apache-2.0. See LICENSE.


Disclaimer

This tool is for authorized security research only. It is designed for:

  • Coordinated vulnerability disclosure in open source projects
  • Security research with responsible disclosure timelines
  • CTF competitions and educational contexts

Do NOT use this tool to:

  • Attack production systems without authorization
  • Exploit vulnerabilities for malicious purposes
  • Bypass responsible disclosure timelines
  • Mass-scan repositories without manual review

The authors assume no liability for misuse. You are responsible for ensuring your research complies with applicable laws and the target project's security policy.