Privacy Firewall for AI Agents

I started exploring multi-agent AI systems because I was fascinated by the idea that instead of asking one AI model to do everything, we could have multiple specialized agents work together, use tools, call APIs and make decisions on a user's behalf. The more I experimented with this idea, the more I started thinking about a problem that felt increasingly important to me. How much of a user's private information should an AI agent actually be allowed to access?

An agent might need to prove that I am eligible for a student discount, for example, but does it really need my student ID, university details or other personal information? An agent helping me make a purchase might need to know whether a payment can be authorized, but should it have access to my entire financial profile? As AI agents become more capable and autonomous, simply giving them access to everything doesn't feel like a sustainable solution.

Around the same time, I had recently started exploring blockchain and became really interested in the privacy side of it. I had initially associated blockchain mostly with transactions and public ledgers, so learning about zero knowledge proofs and the idea of proving that something is true without revealing the information behind it was really interesting to me. This made me think about whether the same concept could be applied to AI agents. That eventually led me to Midnight and the idea of combining privacy preserving blockchain technology with multi agent AI.

That is where Privacy Firewall for AI Agents came from.

The idea is to put a privacy layer between AI agents and the tools or APIs they interact with. Instead of allowing an agent to freely pass a user's information to every service it wants to use, the firewall checks what the agent is asking for, whether it is actually allowed to access that information and whether the request can be satisfied through a privacy preserving verification instead of exposing the underlying data. For example, an agent could verify that a user is eligible for a student discount without unnecessarily exposing their student ID or other credentials. If an agent tries to access something unrelated, such as a user's bank balance, the firewall can block the request.

For me, this project is also the beginning of something much bigger. I am really interested in exploring how this idea could eventually be used outside AI agents as well, especially in areas such as India's KYC systems, government services, digital identity and financial services. There are many situations where people have to repeatedly provide sensitive information simply to prove that they meet a particular requirement. I would like to explore whether privacy preserving verification could allow people to prove that they are eligible for something without having to share everything about themselves every time.

My hope is that privacy preserving technology can eventually become something that is useful in real world systems and can actually be adopted, rather than remaining something that is only interesting from a technical or blockchain perspective. Giving people more control over their personal information while still allowing digital services to verify what they need could make these systems much safer.

This hackathon gave me the opportunity to bring together two areas that I have recently become really interested in, multi agent AI and blockchain based privacy, and experiment with what happens when privacy is treated as a fundamental part of how AI agents operate rather than something added later.