Phishing for Patterns: What Happens When Agents Explore Domain Data
No ratings
Presented at Area41 Conference 2026 by
Over the last five years we’ve been practicing the nerdy hobby of collecting terabytes of valuable domain information and working on a efficient and relatively cheap storage and aggregation system. As this hobby is just about as expensive as diving, skiing or paragliding, we were told to somehow try to cover at least the infrastructure costs. What a brilliant idea! - But since we’re engineers by heart, talking to people is not our strength and procrastination hits hard. With the raise of AI, however, it’s our time to shine. What if we don’t have to talk to people anymore but to AI agents? - They’re a lot more rational, can easily execute the same security analytics use cases and we don’t even have to wear a nice shirt when talking to them. To verify whether this is a legitimate approach, we built an extensive API and an MCP server for our system. Starting with one of the most prevalent use cases in cyber security, finding new phishing sites. We fed our agent real phishing sites, tested multiple different LLMs and asked our agent to find more domains of the same cluster. Furthermore, we requested the findings to be documented in a report or translated into a consumable IoC feed. In this talk we will share technical insights on what we learnt on our journey from an ordinary web UI to an agent focused interface. We will share our experience with agent design, comparing REST APIs vs. MCP servers, and the challenges of guiding agents to consistently choose the right tools for the right tasks. We will also explore context economics and how prompt guidance can both enable and constrain LLM creativity. Last but not least we will examine what LLMs already “know,” and how this differs between powerful models like Opus 4.6 and smaller local models such as Qwen 3.5 or Gemma 4. Finally, we want to close with a discussion on how good current LLMs are with threat hunting given the right data and tools. So far, the AI agent hasn't paid our infrastructure costs. It seems humans aren't the only ones who aren't fond of bills. As engineers, we’re however more than happy to have learnt so much on this journey!