Purpose
DarkMoon is an open-source autonomous AI penetration testing platform for self-owned infrastructure (project address https://github.com/ASCIT31/Dark-Moon）。用户将其指向一个已获授权的目标，它便自主运行完整评估：进行推理、规划，并调度 50 specialized agents execute real offensive operations through a controlled MCP layer, then provide exact commands, raw output, and evidence for each vulnerability. Its official positioning is: it does not replace penetration testers, but rather eliminates the repetitive parts of assessments with evidence, letting experts spend their time on judgment rather than grunt work.
Main Capabilities
- Coverage: 50 specialized agents spanning Web, API, cloud, Active Directory, Kubernetes, CI/CD, and AI/LLM (OWASP LLM Top 10).
- Find and prove: each vulnerability comes with an exact command and raw output (exploitation is asserted by the agent; Pro's remediation retest is a machine-verified step), with almost nothing requiring manual triage.
- Local LLM + privacy gateway: the gateway tokenizes real IPs, hosts, and credentials locally, the model reasons based on placeholders, and real values stay within the user's boundary.
- Open source and auditable: GPLv3, self-hosted, and each agent's methodology is plain Markdown that can be read, diffed, and forked directly.
- Workflow orchestration: agent orchestration dynamically selects and dispatches specialized agents based on the technology stack detected on the target; planes that only make sense with credentials (cloud accounts, CI/CD, secret stores, databases, Active Directory, Kubernetes) are not dispatched based on inference and are only triggered when concrete artifacts (keys, tokens, reachable metadata endpoints) are discovered or the user explicitly authorizes it; otherwise they are flagged in the report.
- Methodology alignment: as a strategic-level AI security agent orchestrator, it aligns with ISO 27001, NIST SP 800-115, and MITRE ATT&CK methodologies.
Automated execution flow: discover the target environment (ports, services, protocols) -> fingerprint the technology stack (frameworks, CMS, API) -> model the attack surface -> deploy specialized sub-agents based on the detected technologies -> execute an intelligent vulnerability scanning loop with reactive adaptive capabilities -> validate findings with evidence (requests, payloads, responses) -> generate a structured audit report. Multiple agents can execute in parallel on a hybrid architecture.
How to Use
The open-source version is a command-line tool (that is what you get after cloning the repository). Start an assessment from the command line and observe the streaming output of each agent, command, and finding in real time. install.sh interactively configures the LLM provider (no need to edit docker-compose.yml) and builds the complete stack, supporting cloudProviders (Anthropic, OpenAI, OpenRouter…) and local models (Ollama, llama.cpp).
Prerequisites: Docker and Docker Compose, plus an LLM API key (or a local model). For specific installation and run commands, GPU settings, environment variables, and the full parameter reference, see the official Full Documentation. Reference links include Quick Start, Install & Run (OSS), Demo (Pro).
Scope definition is supported directly on the command line, including zero-config quick penetration testing and bug bounty mode (the corresponding flags are activated automatically). Key flags include FOCUS, EXCLUDE, CREDS, TOKEN, NOISE, SEVERITY, FORMAT, etc., all naturally interpreted by the AI; EXCLUDE/FOCUS use a free-form syntax and support advanced multi-target scope definition (for details, see Scope Definition in the full documentation).
For integrations, it can run in existing build locations: GitHub, GitLab, Jenkins, VS Code, JetBrains, n8n, Grafana, and Splunk. Integrations labeled OSS + Pro apply to the open-source version (local CLI + JSON); Pro-only integrations consume the DarkMoon Pro REST API. All integrations output only security metadata (severity, status, MITRE, ID), and never evidence, secrets, or tokens. The full guide is at docs.dark-moon.org.
Architecturally: the AI handles reasoning and planning, MCP controls what can be executed, the Toolbox runs isolated tools inside Docker, and the AI never directly touches the system. The built-in tools are purpose-built Docker images; 140+ security tools are compiled and optimized in multi-stage builds, and all tools are ready to use with no path configuration required.
For example prompts and more scenarios (including DVGA, Juice Shop, and headless browser), see Prompt Examples in the full documentation.
Applicable Scenarios
- Security teams: run continuous automated penetration testing against infrastructure.
- DevSecOps pipelines: integrate AI-driven security research into CI/CD.
- Bug bounty hunters: useautonomous target analysis to accelerate ethical hacking work.
- Security researchers: explore the attack surface with an AI cybersecurity platform that can adapt in real time.
- Training and education: learn offensive and defensive security through guided, reproducible assessments.
Limitations and Boundaries
- Should only be used on your own infrastructure or authorized targets.
- The open-source version is a CLI; the web dashboard and automated remediation to pull request are not part of the open-source build.
- The official benchmark is a reproducible black-box run against OWASP Juice Shop (using the cloud frontier model Anthropic Claude), with a total of 57 real vulnerabilities; local model coverage depends on the model being run.
- Surfaces requiring credentials are not dispatched based on inference; they are triggered only when there is a concrete artifact or explicit authorization.
- Paid Darkmoon Pro provides: a web dashboard, findings with CVSS and ATT&CK mapping, exportable web reports (Markdown and PDF), a track-based attack surface map, scheduled recurring runs, an activity view, and automated remediation from discovery to fix (discovery → sandbox-verified fix → human-reviewed pull request → retest against the original exploit, and it never auto-merges). Pro also supports one-click deployment on any cloud (a valid Pro license key is required; no proprietary hardware or cloud marketplace is needed).
- Regarding details such as charges, pricing, and hardware sales: the materials indicate that Pro is a paid version and requires a valid license key, but no specific pricing information is provided.
License: GNU General Public License v3.0. The project is built by ASC-IT and is officially described as open source, AI-driven, and made in France.