IronCurtain – A secure* runtime for autonomous AI agents (github.com)

IronCurtain has been announced as a research prototype designed to create a secure runtime for autonomous AI agents, where security policies are articulated in a human-readable "constitution." This innovative approach addresses the significant concern of ambient authority in existing agent frameworks, where agents have access to users' full permissions. IronCurtain assumes the potential for LLMs (language models) to be compromised and instead emphasizes enforcement mechanisms that operate independently of the model’s behavior, proposing that the agent remains untrusted and security is managed at the boundary. The system allows users to express their security intents in plain English, which IronCurtain compiles into deterministic rules enforced during operation. This avoids the inefficiencies of limiting agents to narrow sandboxes or requiring constant user approvals. Key technical features include the use of a policy engine that mediates all tool interactions to allow, deny, or escalate actions, as well as operating in a V8 isolate to further compartmentalize security. Additionally, IronCurtain facilitates flexible operational modes—such as using built-in agents or external Dockerized agents—making it a versatile solution for enhancing security in AI agent operations while maintaining autonomy and efficiency.