Noisegate

Noisegate is an open-source demonstration of a differential privacy gateway that lets AI agents query sensitive data without exposing individual records. AI agents built on large language models (LLMs) can be wrong, manipulated or deliberately adversarial. If such an agent can query a sensitive database, it might reveal personal information. Noisegate addresses this by ensuring that the privacy guarantee does not depend on the AI being trustworthy. All protections are enforced by trusted code that sits below the model. The developer states that it is a demonstration, not a production product.

Users ask questions about a dataset in plain language, or connect an AI agent through the Model Context Protocol (MCP). The LLM translates each question into a constrained query, rather than free-form database code. A validation layer then checks the query against the dataset's policy: only permitted columns and aggregate statistics are allowed, and queries that single out tiny groups are rejected. A privacy engine runs valid queries and adds calibrated random noise, so no individual's data can be reconstructed. Each answer comes with a confidence interval.

Each user has a limited privacy budget. Every answer uses part of it, and once it is exhausted, further queries are refused. An audit log records each answered query and its cost. To demonstrate its guarantees, the repository includes working versions of three classic privacy attacks, which succeed with privacy disabled and are defeated with it enabled. The noise mechanism is cross-checked against the OpenDP library. Noisegate is written in Python and released under the Apache 2.0 licence.

Auto-discovered on 2026-09-30 by OECD Catalogue Automation