g4mm4p4nd4/VulnHunter ? reverse-engineered prompt

Build me an agentic AI security tool called VulnHunter that helps review source code like an attacker would, not just by matching suspicious patterns. I want it to find real, exploitable issues, trace how an attack could actually work from an entry point to a dangerous sink, and then explain why the issue is valid or discard it if the reasoning does not hold up.

Please make it support a full loop for hunt, fix, and verify, so it can scan code, propose a targeted fix, create a security test that fails first, then confirm the fix blocks the exploit without breaking anything else. It should also have a headless mode for running on multiple repos and opening issues automatically, plus a simple local setup with install and uninstall scripts.

Build it so it works with Claude Code and assumes the user brings their own model access. If you need current docs or best practices, look them up online.