An Isolated, Attested, and Governed Network for Trustworthy AI Research and Deployment: A Reference Architecture, Threat Model, and Roadmap

Abstract— Emerging AI systems increasingly exhibit capabilities—autonomous code generation, tool use, and self-replication potential—that make the public Internet an unsuitable substrate for high-consequence AI research and deployment. This whitepaper reformulates an early-stage dual-system concept (an untrusted “System A” comprising the Internet/IoT, and an isolated “System B” for AI R&D) into a rigorous reference architecture. We contribute: (1) a structured critical assessment of the baseline concept; (2) a layered, zero-trust reference architecture with an explicit trust root, attested cross-domain gateways, confidential tenant workspaces, and least-privilege autonomous guardians; (3) a three-plane protocol redesign (data, control, resource) that corrects the baseline’s conflation of network and resource abstractions; (4) a formalized governance pipeline reconciling ethical screening with auditability, including anonymous-but-linkable telemetry; and (5) a three-phase, metric-gated roadmap traceable to program deliverables (Technical Areas 1.1, 1.2, 2) and aligned with NIST SP 800-207, the NIST AI RMF, IEEE 7000, and MITRE ATLAS. Index Terms— Zero-trust architecture, cross-domain solutions, confidential computing, AI security, trustworthy AI, research networks, provenance, governance.