AIBRIEF
Noisegate, an open source differential privacy gateway, refuses queries that would single out a single record, returns noisy answers to safe ones, and matches OpenDP, the field's reference library, in 35 noise checks.
A privacy gate sits between an AI agent and a sensitive database. When the agent asks a question narrow enough to single out one record, the gate refuses. When the question is broad enough to be safely answered, the gate runs it through a noise machine and returns a useful but imprecise answer. That is the working demo behind Noisegate, an open-source differential-privacy gateway posted to Show HN this week.
Differential privacy is a math framework that adds calibrated noise so no single record can be pulled from the answer. The author's hand-rolled noise routine matches OpenDP, the field's reference library, in all 35 noise-scale checks to within 1e-9. A 250+ test CI suite runs three classic privacy attacks, differencing, membership inference, and singling-out, each succeeding with privacy off and failing with privacy on.
The architecture puts the connecting LLM on the wrong side of the trust boundary. The agent compiles user questions into small constrained queries; a validation layer rejects anything too narrow; a DP engine executes the survivor. The agent never sees a raw row or the privacy budget. MCP, the Model Context Protocol, is the open standard AI agents use to call tools and data sources.
The verification work is real, but the attacks run against the project's own engine. No production deployments, named users, or independent audits appear in the public source. The pattern, a tested validation layer in front of an untrusted agent, is uncontroversial. This specific implementation is the open question.