Human-in-the-loop authorization for AI agents (airlockapp.io)

A new human-in-the-loop authorization system for AI agents, called Airlock, has been announced, enhancing the security and oversight of automated AI actions in enterprise settings. This system allows sensitive commands generated by AI, such as modifying files or executing deployments, to be intercepted and encrypted before human approval. Using AES-256-GCM encryption and Ed25519 signatures, the process requires human oversight at every step, ensuring that no command is executed without explicit authentication and verification. This technology is particularly significant in sectors where AI actions can have substantial impacts, such as DevOps, finance, and security operations. Airlock operates with a robust, cryptographically secure four-step process: the AI agent produces a command, the Enforcer intercepts and encrypts the command, a Mobile Approver reviews it, and finally, the command is executed only upon successful signature verification. The architecture emphasizes security by design, utilizing features like zero-knowledge gateways to prevent exposure of plaintext data and replay protection to guard against unauthorized command executions. Built on the Human Authorization & Review Protocol (HARP), Airlock represents a critical advancement in ensuring that AI operations within enterprises remain auditable and compliant with internal security policies.