Zero Protocol

CAGE: An Autonomous Penetration Testing Copilot Powered by Meta Muse Spark Open The Problem Penetration testers and security researchers spend valuable time on repetitive reconnaissance, attack-surface discovery, API enumeration, and baseline vulnerability testing. They must then investigate noisy scanner output to distinguish genuine security weaknesses from false positives. This leaves less time for complex exploitation and business logic flaws that require human expertise. AI agents offer a way to automate parts of this workflow, but unrestricted execution introduces serious risks. Agents can misinterpret results, repeat ineffective actions, produce unsupported findings, or execute unsafe commands. Effective automation requires not just AI reasoning, but controlled execution and systematic validation. What CAGE Does CAGE (Controlled Autonomous Generation Engine) is an AI-powered penetration testing agent built for offensive security researchers. It automates suitable reconnaissance and common vulnerability-testing tasks, helping researchers move from manual, fragmented workflows toward coordinated, AI-assisted security assessments. Meta Muse Spark Open decides what to investigate next. A deterministic Python engine decides what is allowed to run and what is allowed to be reported. How It Works Reasoning: Meta Muse Spark Open analyzes collected evidence and proposes the next testing steps. Orchestration: An agent loop coordinates tool execution, monitors progress, detects stalled workflows, and manages completion conditions. Controlled Execution: Shell commands pass through command validation and deny-list checks, with execution timeouts. A headless Chromium fallback supports testing JavaScript-rendered web applications when ordinary HTTP requests are insufficient. Finding Validation: A validation layer helps filter likely false positives, including SPA fallback pages, informational noise, and invalid payload results. Persistent Memory: A JSON-based state store and SQLite ledger track testing activity to reduce redundant work and preserve progress across the workflow. A live dashboard provides visibility into agent activity. By separating model-generated decisions from deterministic execution controls, CAGE aims to make autonomous security testing more observable, manageable, and reliable. Why Meta Muse Spark Open? Meta Muse Spark Open serves as CAGE's reasoning engine, connecting an open-weight model to a practical offensive-security workflow. Because the weights are open, the model can be run locally, so sensitive assessment data can stay within the researcher's own environment and under their control. This matters in security work, where client data and findings are highly confidential. CAGE is designed to use long-context reasoning to work with substantial security outputs, endpoint inventories, API specifications, and HTTP traces. The model's role is to guide the testing strategy, while the execution layer remains responsible for enforcing operational controls. Who It Helps CAGE aims to help security researchers, bug bounty hunters, and small security teams spend less time on repetitive testing and more time investigating meaningful vulnerabilities. Findings are filtered by validation before they are reported, and every run can be observed live. Responsible Use CAGE is intended exclusively for authorized security testing in controlled environments or against targets for which explicit permission has been granted.