Australia’s Apate develops and operates AI bots that prolong phone calls and message exchanges with scammers. The bots pose as victims, wasting scammers’ time while collecting information such as scam website links and payment details.
Keeping scammers on the phone to waste their time
The safe response for people is to hang up immediately when a call seems suspicious. Apate’s AI takes the opposite approach. It does not fall for the scam, but shows enough interest to encourage the other person to keep pitching, prolonging the conversation.
The aim is to use up time scammers would otherwise spend on their next target. Unlike measures that protect users by blocking incoming calls, this approach seeks to disrupt the scammers’ work. However, the time spent prolonging conversations does not in itself show how much harm has been prevented.
Founder and CEO Dali Kaafar says it is not unusual for calls with the bots to last more than two hours. The bots have different personalities, language skills, and personas. They do not always answer the phone; sometimes they say they will talk later and hang up. Varying their responses is one way to make them harder to recognize as AI.
Designing conversations that show suspicion but keep listening
The key to the dialogue is not to play someone who believes everything. The bots show suspicion without ending the conversation, leaving scammers room to think they can still persuade them. By moving between wariness and interest, they are designed to get the other person to keep explaining.
When WIRED tried a public demo while playing the role of a scammer, the AI refused to invest, even after being pitched a fictional cryptocurrency investment for about six minutes. The test found that the timing of its responses and the flow of the conversation felt natural. However, this was an experience with a demo, not a test measuring how many real scams it can prevent.
Collecting scam website links and payment details from conversations
Another goal is to collect information. If the other person keeps pitching, the conversation may reveal details used to carry out the scam, such as a URL they direct the target to or bank account details for a transfer. The system does more than waste time: it turns those exchanges into material for investigations.
Kaafar says the system operates about 350,000 bots and has collected more than 250,000 items of information. These include scam website URLs, accounts used to move money, and banking details. These figures were provided by the company and have not been independently verified.
The system’s activities are not limited to phone calls. The bots respond to text messages and are also said to infiltrate online scam chat groups. Banks use the service, and it also receives support from telecom companies. However, the materials provided do not name specific customers or state how many accounts or websites have been shut down using the information collected.
AI that keeps attackers engaged is also being used in decoy systems
Deliberately luring attackers in and observing their behavior is a long-standing cybersecurity defense technique. A common example is a honeypot: a system made to look real so intruders can be studied or kept busy.
Mark Vero, a doctoral researcher at ETH Zurich, and colleagues are studying honeypots equipped with large language models, which handle tasks such as generating text. In the research they presented, AI agents carrying out attacks stayed engaged longer than they did with systems whose responses were predictable, and were less likely to identify the honeypot as a decoy.
While Apate generates human-like conversation, this approach generates responses from systems that appear genuine. Both use AI not simply to detect suspicious behavior, but to engage directly with the other party. However, this research focuses on AI used by attackers and cannot be treated as evidence of the same effect on human scammers.
A record of long calls is not proof of reduced harm
Evaluating this approach requires more than counting conversation time or bots. It is necessary to assess separately how much the scammers’ work was disrupted, whether the information collected led to accounts or websites being shut down, and whether harm was ultimately reduced.
If banks, telecom companies, social media platforms, and law enforcement agencies can share information, clues gathered in conversations could potentially support other countermeasures. However, scammers also use large-scale automated calling and organized operations, so keeping some of them on long calls may not be enough to stop them altogether.
The materials provided show Apate’s operating scale, the number of items of information it has collected, and examples of lengthy conversations. They do not include data on the rate of reduction in financial losses or how much the scam success rate has fallen.
Comments