Noisegate is a differential privacy gateway allowing untrusted LLM agents to query sensitive data over MCP. It provides a formal guarantee that no individual's record can be leaked, even if the agent is adversarial. Enforcement is handled by trusted code below the model, with a runnable attack gallery validating its effectiveness.

The system uses deliberate noise to obscure individual data points while preserving population-scale signals. Three classic privacy attacks—differencing, membership inference, and singling-out—are demonstrated to be defeated when privacy is enabled. The noise mechanism has been independently verified against Open DP, the industry reference implementation, showing high agreement.

Noisegate supports enhanced utility, allowing 3x more queries (308 vs. 100) under the same privacy guarantee through hybrid CDP composition. Built for AI agents, it runs as an MCP server, with the LLM proposing queries that are then validated and executed by a deterministic, fully tested layer. Nothing the model outputs can expand its authority, ensuring a small, secure trust boundary.