Building Trustworthy AI Agents

This essay argues that trustworthy personal AI assistants require decoupling personal data stores from AI models and making integrity the organizing principle of security. It outlines six requirements—broad data ingestion and access across models, provable accuracy, fine-grained user control and auditability, robust defenses against read/write attacks, and ease of use—supported by cryptographic verification, access control, and auditing. The piece references initiatives like the Human Context Protocol and extensions to Tim Berners-Lee’s Solid framework, asserting that only this separation allows security advancements to keep pace independently of AI model performance.