AI penetration testing + local large models + MCP: DarkMoon open-source breakdown, replacing IPs and credentials with placeholders
「AI渗透测试+本地大模型+MCP:DarkMoon开源拆解,把IP和凭据换成占位符」
Jumang Security Lab broke down the GitHub open-source project DarkMoon (ASCIT31/Dark-Moon), an AI penetration testing command-line tool that supports cloud or local LLMs. Its privacy gateway replaces 9 categories of sensitive values, including internal network IPs, hostnames and credentials, with placeholders; the model sees only placeholders, while the true values remain locally as Fernet ciphertext, and MCP provides no restoration tool, so even if prompt injection succeeds there is no legitimate path to swap back the true values; the default policy is DEGRADE, which only degrades and does not refuse. The execution side has three layers of guards: refusing commands that cannot finish before takeoff, enforcing timeout inside the container (default 300 seconds), and giving in-boundary alternatives when refusing. The project publicly ran black-box on OWASP Juice Shop and found 57 vulnerabilities in 28.5 minutes, using Anthropic's cloud model; local model coverage depends on the model used; finding status is given by agent review rather than machine verification, and sensitive values in startup parameters are not yet tokenized beforehand.