Ai on ~Vigneshwar Sundararajan

The Middleman Pipeline

Mon, 08 Jun 2026 00:00:00 +0000

TL;DR: We can protect our private data when using cloud AI by hiding it among highly realistic “fake” data. To keep this fast and cheap, a central “middleman” server groups similar user requests together into one single batch before sending them to the AI.

I spend a lot of time thinking about the gap between artificial intelligence and security. We all want to use powerful cloud AI models to fix our code, summarize medical records, or analyze financial data. The problem is that sending this private data to an external server is a huge privacy risk.

Catching Phishing Sites with AI Teamwork

Wed, 17 Dec 2025 00:00:00 +0000

TL;DR: Single AI models often make mistakes when trying to spot phishing websites. A better approach is using a "multi-agent debate" system. By creating a team of specialized AI agents that debate the evidence together, we can drastically reduce errors and catch complex phishing attacks early. I built a tool called Cross-Check using Google’s Agent Development Kit to bring this concept to life.

As a security researcher, I spend a lot of time looking at how cyber attacks are evolving. Phishing is getting much smarter. Attackers are using AI to create flawless fake websites, so naturally, we want to use AI to defend against them. But there is a problem.

Stop Googling Exploits: Try Mini RAGs

Fri, 19 Sep 2025 00:00:00 +0000

TL;DR: I built Kestrel, a simple command-line tool using RAG and ChromaDB, but the real goal is much bigger. This project highlights the power of mini, lightweight RAGs. Imagine a future where we do not need to read man pages or memorize complex filters—we just send one prompt and get our freedom back.

The Annoying Problem

When you are deep into cybersecurity research or testing, finding the right exploit or command can be frustrating. You spot a vulnerability, but then you have to stop what you are doing. You open Exploit-DB, run slow and complex search filters in Metasploit, or read through massive man pages just to find the right flag.