Foresight Institute

Andrew Trask | It’s Time to Harvest the Secure AI Tech Tree

Andrew Trask opened the session by presenting the "secure AI tech tree," a visual framework that maps the sprawling landscape where cryptography, deep learning, and distributed systems converge. The tree groups roughly five major subdivisions—privacy‑preserving collaboration, attribution control, trust mechanisms, and related sub‑clusters—each populated with specific research problems and emerging solutions. He argued that many of these disparate issues trace back to a single, under‑appreciated root: insufficient attribution‑based control, which he broke into three sub‑problems—addition, copy, and branching. In practice, this manifests in federated learning deployments, where gradient updates can leak private data unless both input and output privacy (e.g., differential privacy) are applied. Trask highlighted the need for hierarchical trust‑over‑IP models, suggesting that aggregating updates within trusted sub‑networks before broader dissemination can curb poisoning and copy attacks. Concrete examples punctuated his talk: Google’s 2017‑18 federated learning rollout for on‑device language models, the vulnerability of gradient leakage without differential privacy, and the paradox that a model trained on a fraction of the world’s data can still expose individuals when combined with output‑privacy gaps. He also referenced collaborative work with William Isaac, Eric Drexler, and others on privacy‑enhanced technologies that map specific neural activations to data sources, illustrating how attribution can be enforced. The broader implication is that by collapsing a sprawling list of problems into a tighter set centered on attribution control, researchers and firms can prioritize tooling, standards, and cross‑disciplinary partnerships. This streamlined focus promises faster advances toward AI safety, regulatory compliance, and commercial trust, turning abstract security concerns into actionable product roadmaps.