Honeypot agent

A honeypot agent is the agentic equivalent of a classic security honeypot: instead of a fake server left open to attract attackers, it is an AI agent instructed to play a convincing target. The agent engages a scammer, fraudster, or attacker long enough to gather intelligence that a human researcher could not realistically collect at scale. Thousands of agents can run simultaneously, each holding a different conversation.

Australian startup Apate.ai is the most visible current example. The company deploys fleets of agents that pose as innocent fraud victims across calls, messages, email, and web forums. When a scammer engages, the agent keeps them talking while surfacing bank accounts, wallet addresses, malicious domains, and other operational details. That data goes to customers in telecom, banking, and government who use it to block infrastructure or assist law enforcement.

The pattern is expanding beyond fraud. Security researchers use similar approaches to map phishing kits, study social engineering scripts, and test whether AI-generated content can reliably pass as human to adversarial models. The main risk is that the agent itself becomes a target for prompt injection: if an attacker realizes they are talking to a bot, they may attempt to manipulate it into revealing system instructions or taking unintended actions.