AI systems are being deployed to waste scammers’ time and gather intelligence

Anti-cybercrime initiatives are increasingly using AI systems designed to engage and mislead scammers, wasting their time and gathering actionable intelligence on fraud operations. These tools deploy lifelike AI bots that scammers believe are real victims, prolonging interactions and diverting resources away from actual targets.

The Australian company Apate has built a platform that diverts phone scammers into extended conversations with AI bots trained to simulate vulnerable victims. The bots are engineered to remain engaged without falling for the scam, while collecting data such as scam URLs, money mule accounts, and bank details. According to Apate’s founder and CEO Dali Kaafar, the platform operates around 350,000 bots that not only intercept calls but also infiltrate scam chat groups online and respond to text messages with the same goals of frustrating scammers and gathering intelligence.

Kaafar describes the bots as designed to be "the perfect victims," noting that even a single minute spent engaging a scammer could prevent hundreds or thousands of potential victims from being targeted. The bots vary in personality, language skills, and profiles to avoid detection, and some are programmed to hang up or delay responses to mimic human behavior. Apate’s platform is used by banks and supported by telecom companies, and the company claims to have collected over 250,000 pieces of real-time information from fraudsters.

Testing by WIRED’s Kernel Panic newsletter using a demo version of Apate’s system found the AI bots responsive and capable of sustaining natural-sounding conversations, even when challenged by multiple human participants role-playing as scammers. The testers were unable to trick the bot into engaging with a cryptocurrency scam within six minutes, highlighting the system’s resilience to manipulation.

Beyond Apate’s efforts, researchers are exploring additional AI-driven tactics to disrupt cybercriminals. At ETH Zurich, doctoral researcher Mark Vero and colleagues have incorporated large language models (LLMs) into honeypot systems—decoy environments designed to attract hackers. Their research indicates that LLM-powered honeypots keep AI-driven attackers engaged significantly longer than traditional honeypots, with attackers less likely to identify the decoy as a trap. Vero states that well-designed LLM honeypots could provide a substantial advantage for defenders by prolonging attacker engagement and improving intelligence collection.