The Pulse
Apate’s 350,000 Bots Keep Scammers Talking
Australian company Apate uses AI personas to occupy phone, text and chat scammers while collecting information about their operations. A separate ETH Zurich study finds language-model honeypots keep AI hacking agents engaged longer than rul
AI.info Team · Written with AI ·
In a demo reported by WIRED on October 10, two journalists spent six minutes trying to persuade an AI persona to invest in a cryptocurrency pitch. The bot stayed engaged but did not take the bait. The test offered a small-scale view of a new defensive tactic: give scammers convincing targets that waste their time instead of rewarding their scripts.
Apate builds a crowd of artificial targets
For about two years, Australian company Apate has developed bots that answer scam calls and hold conversations without falling for the fraud. The company’s founder and CEO, Dali Kaafar, describes the aim this way: “we’re building the perfect victims for scammers.” The bots are meant to keep operators occupied while gathering details about the schemes they run.
Kaafar told WIRED that Apate’s platform has around 350,000 bots and is used by banks with support from telecommunications companies. The bots also respond to texts and enter scam chat groups. Kaafar says the company has collected more than 250,000 pieces of information about fraudsters, including scam URLs, money-mule accounts and bank details.
The personas vary in profile, language and behavior. Some may appear to have WhatsApp; others may end a call and say they will return later. In the journalists’ demo, a bot called Lucy maintained a conversation while the testers also played a friend and a financial adviser. Kaafar says some Apate calls last more than two hours.
Time spent talking is not the same as losses prevented
The approach aims to exploit scammers’ reliance on volume. If a caller is tied up with a bot, Kaafar argues, that is time the caller cannot spend reaching other people. He told WIRED that a minute spent talking to one of the bots could mean “saving hundreds, if not thousands” of potential targets from that same scammer. That is Kaafar’s estimate, not a measured count of people protected.
The report describes the system’s scale and intelligence collection, but does not establish how many scams Apate has stopped, how much money it has recovered or whether bot conversations reduce victimization overall. And while a prolonged exchange may consume a scammer’s time, the tactic does not stop the wider flow of calls and messages. WIRED notes that cybercriminals send billions of calls and messages each year.
ETH Zurich tests decoys against AI attackers
A separate line of work applies language models to digital honeypots: decoy systems designed to attract attackers and reveal their methods. In a 2026 study, researchers evaluated AI-powered HTTP honeypots against hacking agents using 16 simulated backend applications. The results concern web-system attacks, not phone scams, but test a related idea: can a realistic artificial target keep an attacker busy?
“they lead to substantially longer interactions with the attacker than rule-based baseline honeypots”
The authors of “Honeyval: A Comprehensive Evaluation Framework for LLM-powered HTTP Honeypots”
The paper reports an average of 82.6 HTTP requests per interaction with language-model honeypots, compared with 30.6 for rule-based versions. In most cases, the agents identified the decoys as honeypots at rates below 40 percent. Those are controlled evaluations of simulated web applications; they do not show that scammers on phone lines are being diverted at the same rate.
Disruption depends on what happens after the conversation
Apate’s system and the research honeypots turn conversational AI toward deception in a defensive setting. Their immediate output is time and information: a scammer kept talking, or an attacker’s requests expose how it probes a fake system. Converting those signals into action requires organizations to identify the people, accounts and infrastructure behind a campaign and share useful findings with the institutions that can respond.
WIRED points to the need for better intelligence sharing among law enforcement, banks, social media companies and other industries. The reported figures show that Apate has built a sizable network of artificial targets and gathered a large pool of fraud-related information. They do not, on their own, show how much that network has reduced fraud.