Apate’s 350,000 AI bots keep scammers talking while they gather data

The cost of keeping a scammer talking

Australian company Apate has spent the past two years building a system that routes scammers’ calls to AI bots. The bots try to keep callers engaged without falling for their pitches, giving criminals reason to stay on the line while Apate gathers information about their methods. Banks use the platform, and telecommunications companies support its operation.

Founder and CEO Daly Kafaar says the system has about 350,000 bots. They answer calls, enter scammers’ group chats and reply to text messages. Apate has collected more than 250,000 pieces of information in real time, including links to scam websites, details about money mules and bank account information.

The bots vary their personalities, language skills and habits to seem like different people. Some use WhatsApp; others do not. Some answer calls, while others hang up and say they will get back to the caller later.

A convincing target, not a convincing promise

Journalists at Kernel Panic tested an Apate demo by posing as scammers and talking to one of its AI characters. The bot was skeptical enough to resist, but not so skeptical that it ended the exchange. The testers found the interaction irritating—a reaction that, in this case, suggests the bot was doing its job.

Kafaar says Apate calls often last more than two hours. In the demo, the bot stayed on topic and kept up a natural pace. The journalists took turns playing Lucy, a friend, and Mickey, her financial adviser. After six minutes of trying to persuade the AI to invest in cryptocurrency with supposedly incredible returns, they had not succeeded.

The scale of the operation is striking, but it does not mean ordinary people will stop getting scam calls or messages. Cybercriminals send billions of messages and make billions of calls each year. The most sophisticated operations run from large industrial-scale scam compounds, and years of work by investigators and people infiltrating criminal networks have not stopped digital fraud from growing.

The broader contest is about attention

AI could give defenders another way to disrupt scammers: keep them busy, learn how they operate, and make it harder for them to pursue other targets. That fits a wider push for law enforcement, social networks, banks and other industries to share information more effectively, with AI helping monitor and analyze it. Some initiatives have also tried to exploit criminals’ psychological weaknesses, including direct efforts to provoke participants in ransomware attacks.

I think the key measure is not how many bots Apate can field, but how much criminal effort they divert—and whether the information gathered can help others act. The announcement offers evidence of a large network and a substantial collection of scammer data, but no measure of how much fraud either has prevented.

AI is also changing the other side of the deception. Security researchers have long used honeypots—fake virtual machines designed to attract hackers, waste their time and reveal their methods. Mark Vero, a computer science PhD student at ETH Zurich, says developers of open-source honeypots are increasingly adding large language models to make them more plausible.

In recent research by Vero and colleagues, AI agents attacking a honeypot spent significantly longer on one with a language model than on one with more predictable behavior. The agents were also more likely to mistake the model-based imitation for a real system and less likely to recognize the trap. Vero says well-designed systems could give defenders an advantage.

The tension is clear: a more convincing decoy can consume an attacker’s time, but its value depends on what defenders learn and do with that time. Better imitation alone does not stop the billions of calls and messages still reaching potential victims.