🛠 IronCurtain — A Secure Runtime Environment for Autonomous AI Agents
The IronCurtain project allows for describing security rules in natural language in the form of a "constitution." These rules are compiled into deterministic policies that are verified in real-time during tool calls via the MCP protocol.
🌍 It solves the problem of "ambient authority" (excessive privileges) by limiting the agent's capabilities at the execution level, rather than just at the instruction level.
👤 It enables the use of powerful agents (such as Claude Code or Goose) with reduced risk that errors or prompt injections could lead to data theft or file deletion.
Source 1: https://github.com/provos/ironcurtain