What is redforge-ai?

RedForge AI is a Python tool for testing the security of AI systems such as chatbots, retrieval based question answering systems, and AI agents that can use tools. It is meant for authorized security testing only, and the project is explicit that it is not a hacking framework or a generic scanner, but a controlled way to evaluate systems you already have permission to test. The tool runs what it calls scoped campaigns: you define exactly which target to attack, which hosts are allowed, and how many attack attempts the run is allowed to make. During a run it records detailed evidence, including the exact payloads sent, the model's responses, any documents retrieved, tool calls made, and any changes to the system's memory. After the run, it produces a report in Markdown and HTML that a developer or security reviewer can read and act on. It looks for problems specific to AI systems, such as prompt injection, jailbreak attempts, poisoned retrieval documents, misuse of tools, and unwanted information leakage. You can run it from the command line, or start it as a web service using FastAPI so other tools can trigger campaigns through an API. It can target a demo agent that comes built in, a general web endpoint, or an OpenAI-compatible API such as OpenAI, Anthropic, or Gemini. It also supports testing systems made of multiple AI agents working together, tracking how tasks get handed off between them and where trust boundaries are crossed. To try it, you need Python 3.11 or newer and the uv package manager. Cloning the repository and running the included setup and demo commands will run a sample security test against the built in vulnerable demo agent and generate a sample report you can inspect. The project describes itself as an early preview, with an architecture split into separate packages for the core engine, plugins, attack packs, target adapters, and the API and CLI layers.