Google AI Updates: October 3, 2026

1. Google Research Moved Federated Learning Into Trusted Execution Environments for Verifiable Privacy

Google Research. Google described a next-generation federated learning system that encrypts data on the device and only lets it be decrypted inside Trusted Execution Environments (TEEs) running pre-authorized workloads, with the access policies published to the Rekor transparency log so outside parties can verify them. A key management service replicated across TEE clusters with RAFT consensus controls access to decryption keys, and operators see only metrics and differentially private model weights. Because more of the computation now runs on servers instead of phones, Gboard has adopted the system for its English and Japanese next-word prediction models, gaining stronger privacy guarantees with smaller noise multipliers and training runs that used to take one to two months now finish much faster. The code is open source in the Confidential Federated Compute repository with reproducibly buildable binaries, and the paper is on arXiv. Source