Automating Phishing Infrastructure Development Using AI Agents

Fred Heiding and Simon Lermen present research on using AI agents to automate phishing from start to finish, not just to draft emails. They recap an earlier study in which AI-written emails performed about as well as those from human experts, then describe newer work where an agent gathers public information about a target, writes a personalized message, and adjusts its persuasion style over repeated attempts. The work extends to voice phishing, and a live demo showed an AI caller asking a volunteer for an email address.

The middle of the talk covers the supporting infrastructure. The speakers describe an agent that searches domain registrar APIs, compares prices, favours cheap domain endings, buys a domain, configures DNS and certificates, and deploys a cloned or generated website. They note that some models could be steered past safety filters with simple prompts, and that cloning a site from screenshots is easy. Cost figures were given for a one-minute AI voice call, around ten cents, to show how cheap sophisticated attacks are becoming.

For defenders, the speakers suggest priming AI classifiers to look for suspicion, which they say raised true detections without increasing false positives. They also propose personalized spam filters, stricter know-your-customer checks for domain registration, and better protection for older adults, who are heavily targeted by scams. The takeaway is that the cost of convincing social engineering is falling, so defensive budgets and research need to keep pace.