July 30, 2026
Private data, public chaos
Show HN: Noisegate – a differential-privacy gateway for untrusted AI agents
AI gets the keys to private data, and the comments are screaming both genius and nightmare
TLDR: Noisegate is a new tool that lets AI query sensitive databases while blocking attempts to expose any one person, even if the AI goes rogue. Commenters loved the “assume the bot is shady” approach, but argued hard over whether it’s a real privacy breakthrough or just a slick refusal machine.
Hacker News got served a classic "this is either the future or a disaster" moment with Noisegate: a tool that lets an AI ask questions about sensitive data while adding enough fuzziness that, in theory, no single person can be exposed. The creator’s big flex is simple even for non-experts: the AI is treated like an untrusted intern. It can ask for totals and trends, but if it tries to get too nosy, the system either adds so much noise the answer is useless or just flat-out refuses. Commenters were obsessed with that dynamic. Half the room was basically yelling, "finally, someone built security for AI that assumes the AI will misbehave". The other half immediately went into detective mode, asking whether this is truly safe or just a very polished demo.
The spiciest reactions came from the project’s attack gallery, where the system shows common privacy tricks working when protections are off and failing when they’re on. That won over a lot of skeptical readers who are tired of security promises with zero proof. But there was still drama: some joked that the product is really a "fancy machine for saying no", while others argued that noisy answers and query limits are exactly what real-world health and payroll systems need. The comedy practically wrote itself, with readers comparing the AI to a casino addict burning through its question budget trying to beat the house. The vibe? Equal parts applause, side-eye, and nerdy popcorn.
Key Points
- •Noisegate is presented as a differential-privacy gateway that lets untrusted AI agents query sensitive data without exposing individual records.
- •The article demonstrates enforcement through noisy answers, query-budget exhaustion, and rejection of overly narrow queries at the trust boundary.
- •Three classic privacy attacks—differencing, membership inference, and singling-out by re-identification—are included as working tests that reportedly succeed with privacy off and fail with privacy on.
- •The project states that its noise mechanism was checked against OpenDP and that large-sample distribution tests passed.
- •The system treats the LLM as untrusted, enforcing privacy in a downstream validation and execution layer, and is implemented with Python, DuckDB, FastAPI, Streamlit, MCP SDK, Docker, and GitHub Actions.