What is agentforge-redteam?

AgentForge Adversarial AI Security Platform is a tool that automatically tests the safety and security of an AI system used in clinical healthcare workflows, specifically an AI "co-pilot" that helps with patient charts and clinical decisions. The problem it solves is that manual security testing (hiring a human team to probe a system for vulnerabilities) cannot keep up with the pace at which AI tools are updated, and upcoming healthcare regulations require continuous, documented risk analysis of any AI tool used with patient data. The platform works by running four AI agents that each play a different role. A Red Team agent generates attack prompts and tries to find ways to make the clinical AI misbehave, for example, causing it to leak one patient's information to another, fabricate citations, or perform unauthorised actions. A Judge agent scores each attack attempt against a written rubric to decide how serious it is. An Orchestrator agent decides which attacks to try next based on what has been covered and what budget remains. A Documentation agent writes up findings and files them, automatically for lower-severity issues, but routing high-severity ones through a human approval queue first. The whole pipeline is coordinated by a LangGraph state machine (a framework for building multi-step AI workflows) and stores its audit trail in a database. The platform targets six threat categories including prompt injection, data exfiltration, and tool misuse. It ships with an operator command-line interface for starting sessions, halting the system, and reviewing the approval queue. The full README is longer than what was provided.