Show HN: Noisegate – a differential-privacy gateway for untrusted AI agents (github.com)

Noisegate, a new differential-privacy gateway designed for untrusted AI agents, has been announced, providing users with access to sensitive data while ensuring that no individual's record can be leaked, even when dealing with adversarial AI. The system employs a unique mechanism where an AI agent generates queries that are then validated and executed through a privacy engine, which adds calibrated noise to ensure strong privacy guarantees. This approach allows for greater utility by enabling up to three times the number of permissible queries when compared to naive budget accounting methods, all while preserving robust privacy for individual records. This innovation is significant for the AI/ML community as it demonstrates a reliable method for enforcing privacy protections in AI applications, without needing to trust the AI agent itself. The implementation aggressively tests its own resistance against classic privacy attacks—such as differencing and membership inference—and employs verifiable mathematical techniques that match the industry-standard OpenDP implementation closely. By employing a rigorous validation layer and tracking cumulative privacy loss, Noisegate ensures that even sophisticated querying does not compromise individual privacy, marking a potential breakthrough in the practical applications of differential privacy in AI systems.