Apate's AI bots divert phone scammers into calls with fake victims

WIRED's weekly Kernel Panic! newsletter, written by Lily Hay Newman and Matt Burgess, argues that AI is no longer only a tool for scammers: it is also being recruited against them. Governments have struggled to deal with online crime that originates beyond their borders, which has encouraged alternative approaches such as using automation to go after scammers at scale, perhaps by spamming the spammers to squander their resources. The authors say AI platforms are now turning those experiments into real tools.

The main example is Apate, an Australian company named after the Greek goddess of deception. For the last two years it has been building a system that diverts phone scammers onto calls with AI bots. The bots are trained to keep conversations going as long as possible without ever actually falling for the scam. The goal is to give scammers enough hope that they stay on the line, while the system collects intelligence about the scams they are running.

Dali Kaafar, Apate's founder and CEO, puts it this way: "What we really like to think is that we’re building the perfect victims for scammers." He says a minute a scammer spends talking to a bot is a minute in which hundreds, if not thousands, of other possible targets are not being contacted by that same scammer, because scammers use automated dialing tools.

According to Kaafar, Apate's platform, which is used by banks and supported by telecom companies, has around 350,000 bots. They do not only pick up calls: they also infiltrate scam chat groups online and respond to text messages. He says the company has collected more than 250,000 pieces of information about fraudsters in real time, from scam URLs to money mule accounts and bank details. The bots have different personalities, language skills and profiles so that scammers hopefully do not detect an AI. Like real people, he says, sometimes they have WhatsApp and sometimes they do not; sometimes they pick up the phone and sometimes they hang up saying they will come back later. Kaafar says Apate calls regularly go on for more than two hours.

The Kernel Panic authors tested a demo version in which the user plays the scammer and talks to one of Apate's AI victims. The personas are designed to show a healthy amount of skepticism while leaving enough openings for the scammer to keep trying, a dynamic the authors say they felt immediately and found intensely frustrating. The system was engaged and responsive and the timing felt natural, even with both authors tag-teaming the AI persona "Lucy" and her financial adviser "Mickey". After six minutes of effort they did not get it to invest in their cryptocurrency opportunity.

The authors keep the scale in perspective. Cybercriminals initiate billions of messages and calls each year, the most sophisticated operations run industrial-scale scamming sites, and law enforcement and research efforts have not been able to stop the overall expansion of digital scamming. They say better intelligence sharing between police, social media companies, banks and other industries is needed and could potentially be aided by AI-driven monitoring and data analysis. Some efforts also try to exploit cybercriminals' psychological vulnerabilities, in some cases by directly trolling those behind ransomware attacks. Using generative AI to keep criminals preoccupied, wasting resources and not launching other attacks, appears to be increasingly effective, they write.

The second example is honeypots, the false virtual machines that companies and security researchers have long deployed to attract hackers, waste their time and gather information about their techniques. Mark Vero, a doctoral researcher at the department of computer science at ETH Zurich, says open source honeypot providers have increasingly been putting LLMs into their systems to make them look more realistic. In recent research, Vero and colleagues found that an LLM-powered honeypot kept AI agents attacking their system "significantly longer" than a honeypot with more predictable behavior. In Vero's words, the agentic attackers are much more convinced by the LLM-simulated honeypots and mark them as actual honeypots at a much lower rate. If such systems are built well enough, he says, they are quite advantageous for defenders.

Key facts

Apate, an Australian company named after the Greek goddess of deception, has spent two years building a system that diverts phone scammers onto calls with AI bots trained to keep them talking without ever falling for the scam.
Founder and CEO Dali Kaafar says the platform has around 350,000 bots, has collected more than 250,000 pieces of information about fraudsters in real time, and sees calls that regularly run past two hours.
The bots also infiltrate online scam chat groups and answer text messages, with varied personalities, languages and profiles.
In the Kernel Panic authors' demo test, the AI persona was engaged and natural, and it did not invest in their fake crypto scheme after six minutes.
ETH Zurich researcher Mark Vero and colleagues found an LLM-powered honeypot kept AI attacking agents engaged "significantly longer" than a more predictable honeypot.

Why it matters

Law enforcement and research efforts have not been able to stop the overall growth of digital scamming, and cybercriminals initiate billions of messages and calls each year. The article describes AI being used to turn the idea of spamming the spammers into a working tool: each minute a scammer spends with a bot is a minute not spent on real targets, and the calls yield intelligence such as scam URLs, money mule accounts and bank details. The same logic shows up on the technical side, where LLM-driven honeypots hold AI attackers longer.

Who it affects

Potential scam victims are the intended beneficiaries, since scammers' time is tied up. Apate's platform, per Kaafar, is used by banks and supported by telecom companies. Security teams that run honeypots are also affected, because open source honeypot providers have increasingly been adding LLMs to make their decoys look more realistic. Scammers and attackers, human or AI, face targets that are harder to tell from real ones.

How to use it

The article does not present Apate as something readers can switch on themselves; it says the platform is used by banks and supported by telecom companies. No pricing, launch dates or countries of deployment for Apate are stated. The Kernel Panic authors tried a demo version in which the user plays the scammer and talks to an AI victim. For defenders, the practical pointer is the honeypot work: Vero says open source honeypot providers are already incorporating LLMs.

How solid is it

Mixed. The Apate figures (around 350,000 bots, more than 250,000 pieces of information, calls over two hours) all come from Kaafar, and no independent verification is given. The authors' own test was a six-minute demo in which they played the scammer, not a live deployment. The honeypot research is described through Vero's comments: the article does not name the study, venue or date, and gives no numeric size for the "significantly longer" effect.

Risks and caveats

Even the authors note that, with Apate's swarm of bots on the case, it is only a matter of time before you get your next scam call or text. No data is given on how many scam calls are actually diverted or on any measured reduction in scam victims. No banks or telecom companies are named, and the article does not say how the bots are built or which models they use. The authors describe the approach as appearing to be increasingly effective, which is their assessment rather than a measured result.

“What we really like to think is that we’re building the perfect victims for scammers”

— Dali Kaafar, founder and CEO of Apate