What would it take for AI agents to work with sensitive enterprise data without weakening trust, auditability, or policy control? Hypha is a proposed secure data network for enterprise AI that studies how autonomous agents may access distributed datasets under explicit delegation, auditable access controls, and policy-constrained execution. This talk presents the system design and research agenda behind Hypha, with a focus on two core problems: private metadata query interfaces that let agents evaluate dataset relevance without exposing raw data, and delegated access mechanisms that support time-bound, compliant data use in workflows such as private retrieval-augmented generation. The broader goal is to develop infrastructure for multi-agent systems that operate over regulated enterprise data with stronger guarantees of accountability and governance.

1. Secure AI data infrastructure for enterprises:
   The project proposes building Secure Data Enclaves for Enterprise AI to allow financial institutions and enterprises to safely use AI agents while meeting strict data security and compliance requirements. It focuses on protecting sensitive data access, provenance, and usage in private RAG (Retrieval-Augmented Generation) systems.
2. Privacy-preserving data sharing and AI workflows:
   The solution combines technologies such as Trusted Execution Environments (TEE), Multi-Party Computation (MPC), Zero-Knowledge Proofs (ZKP), and blockchain-based data networks to enable secure data storage, encrypted sharing, and private metadata queries without exposing actual datasets.
3. Development of enterprise-ready AI security tools:
   The project aims to create MVPs including a private metadata query system and a secure data access pipeline that can integrate with AI agent frameworks. The long-term goal is to provide APIs or protocols enabling compliant AI collaboration across organisations with proper access control and audit trails.