This AI Agent Is Designed to Not Go Rogue

IronCurtain is an open-source research prototype by Niels Provos that isolates agentic AI in a virtual machine and uses a large language model to translate plain-English user policies into enforceable security rules that mediate the agent's access to accounts and services. Designed to prevent harmful or unintended agent behaviors (such as deleting emails or launching phishing), the system is model-independent, maintains audit logs, and refines user "constitutions" over time through edge-case prompts and human feedback; it is presented as a research project seeking community contribution rather than a consumer product.