Noisegate: a differential-privacy gateway for untrusted AI agents
Summary
Noisegate is a differential-privacy gateway that lets an untrusted LLM agent query sensitive data under a formal privacy guarantee. The project demonstrates a constrained NL→AST compiler, a trusted guard/validation layer, and a DP engine with a budget, validated by an attack gallery showing that privacy holds even when the model is adversarial. It includes cross-checks with OpenDP and multiple deployment modes for local testing and AI-agent integration.