Capital One's VulnHunter Fork Breaks Free From Claude to Hunt Any Codebase

A community fork of Capital One’s VulnHunter removes the original project’s Claude Code-specific installation paths and model checks. Maintainer nealbridges preserved the attacker-first scanning workflow and added environment-based configuration so teams can install its skills and agents into another compatible coding harness. Developers can now evaluate the method without adopting Anthropic’s CLI as their primary environment.

Capital One’s original announcement described an Apache 2.0 security tool that applies attacker-perspective analysis to source code. An agentic scanner gives a model a sequence of tasks and tools, allowing it to investigate a suspected flaw across multiple steps. VulnHunter starts at externally influenced entry points, including API routes and file uploads, and traces whether controlled data can reach a dangerous operation, often called a sink. The upstream project ran an Anthropic Claude Opus model inside Claude Code and identified broader model and harness support as a future direction.

Proof before alerts

VulnHunter moves each candidate vulnerability through a falsification stage designed to uncover faulty assumptions, missing links in the attack path, and controls that would stop exploitation. Findings that survive must include an executable proof of concept and a proposed fix. This workflow focuses review time on reachable defects while retaining evidence that developers can reproduce.

The fork packages that process into three primary skills:

• /vulnhunt performs reconnaissance, launches parallel investigations, challenges candidate findings, and checks whether the required attacker capabilities fit the threat model. Its output contains verified findings with executable exploits and proposed fixes.
• /vulnhunter-fix creates an exploit demonstration, writes a failing security test, implements the remediation, confirms that the test passes and the exploit fails, and opens a pull request.
• /vulnhunt-fix-verify uses a separate read-only agent to assess the remediation independently, preventing the fixing agent from grading its own work.