TrustedARI Enhances Trust and Privacy in AI Agent Routing

AI agents frequently interact with various external models, tools, and services through Agentic Routing Infrastructure (ARI). While ARI simplifies managing diverse interfaces and subscriptions, it introduces significant security vulnerabilities. Specifically, ARI typically gains plaintext access to agent queries and service responses, making it difficult for agents to verify service provider authenticity or ensure data integrity. TrustedARI is proposed as a solution to these trust issues, offering a trust-native architecture. It incorporates three key innovations: an ARI-adapted three-party TLS handshake for joint authentication of service providers, a privacy-preserving protocol for collaborative query construction, and a verifiable billing system that maintains data confidentiality. Experimental evaluations demonstrate TrustedARI's efficiency. The adapted handshake reduces communication overhead by nearly 40%, while the privacy-preserving query protocol adds minimal computational and communication costs. The verifiable billing protocol also significantly speeds up proof generation, and the system is designed for easy deployment without requiring modifications to existing service providers.