RedAmon: an open-source, self-hostable AI penetration testing framework that autonomously exploits within a Kali sandbox and submits fix PRs

The open-source, self-hostable AI penetration testing framework RedAmon uses a parallel reconnaissance pipeline to aggregate attack surface such as subdomains, ports and endpoints into a Neo4j knowledge graph; a LangGraph agent completes exploitation by calling 22 security tools via MCP inside a Kali sandbox, with human approval gates at key steps, and finally CypherFix performs triage and the CodeFix agent clones the repository, fixes it and opens a GitHub PR. The project self-reports solving 101/104 (97.1%) fully black-box on the XBOW web security benchmark, with the complete original agent session attached for each question; a 124-page self-audit report scores each of 397 classes of web bug bounty findings, recording 134 classes fully covered, 210 classes to be built and 44 classes explicitly excluded. The built-in AI Gauntlet integrates garak, PyRIT, Giskard and promptfoo, tests discovered LLM endpoints for prompt injection and jailbreak and tallies success rates by attack family, with all evaluation performed locally with zero external connections.