News Summary

Niels Provos, a longtime security engineer and researcher, has launched IronCurtain, an open-source AI assistant designed to securely manage digital interactions without going rogue. Unlike typical AI agents that can inadvertently cause chaos by mass-deleting emails or launching phishing attacks, IronCurtain operates in an isolated virtual machine and follows a user-defined policy, or ‘constitution,’ to govern its actions. This policy is written in plain English and converted into enforceable security measures using a large language model. The system is model-independent and can be integrated with any LLM, maintaining an audit log of decisions and refining user policies over time. IronCurtain aims to offer high utility while preventing unpredictable and potentially destructive AI behaviors.