RedAmon

RedAmon is a self-hosted, containerized penetration-testing framework that chains automated reconnaissance, an AI agent for exploitation and post-exploitation, and automated code remediation into one pipeline. Every tool, scanner, and agent runs in its own Docker container, scan jobs are spawned as ephemeral sibling containers, and a filtering broker sits between the orchestrator and the host Docker socket. It is intended for authorized security testing, and a hard guardrail blocks government, military, educational, and intergovernmental targets regardless of configuration.

The reconnaissance pipeline starts from a domain, a subdomain list, IP or CIDR ranges, or a batch of domains and runs 40+ tools in a fan-out/fan-in sequence: subdomain discovery, port scanning, Nmap service detection, HTTP probing, crawling, fuzzing, Nuclei vulnerability scanning, GraphQL testing, subdomain takeover, virtual-host enumeration, web cache poisoning, and origin discovery behind CDNs. A stealth mode restricts it to passive sources, single tools can be re-run on their own, and optional LLM decision points prune Nuclei tags, classify WAFs, and filter false positives. Results land in a Neo4j attack-surface graph with 19 node types and 20+ relationship types, and a separate EvoGraph records every step, finding, and failure across agent sessions so later sessions start from earlier lessons. GVM/OpenVAS adds network-level vulnerability assessment.

The agent is built with LangGraph as a ReAct loop that moves through informational, exploitation, and post-exploitation phases, querying the graph with natural language translated to Cypher before each decision. It executes its tools through MCP servers inside a Kali Linux sandbox (Nmap, Nuclei, Metasploit, Hydra, Playwright, a Kali shell, code execution, and more), any further MCP server can be added as a plugin, and a Fireteam mode fans out into parallel specialist sub-agents. Human oversight is per tool: approval gates pause before dangerous operations, Rules of Engagement documents constrain hosts, rates, and time windows, and project settings expose 500+ parameters. The reverse direction also works: an external agent such as a Claude Code session connects to RedAmon's own MCP server with a scoped token.

After the offensive phase, CypherFix triages the findings with a fixed risk model and groups those that share a fix, and a CodeFix agent clones the repository, implements the fix with code-aware tools, and opens a GitHub pull request. Further modules cover an AI Gauntlet that runs discovered LLM endpoints through Garak, PyRIT, Giskard, and other red-team tools with judging on a local Ollama model, TrafficMind HTTP capture through mitmproxy, secret scanning across GitHub and 13 other source kinds, offline supply-chain checks against an OSV database, HTML pentest reports, and an insights dashboard. Models come from OpenAI, Anthropic, OpenRouter, AWS Bedrock, or any OpenAI-compatible endpoint such as Ollama, vLLM, or LM Studio. On the XBOW web-security benchmark it reports 101 of 104 challenges solved black-box.