Artificial intelligence is giving fraud fighters a new kind of defensive tool: a victim who can never lose money. WIRED reporters Lily Hay Newman and Matt Burgess describe a growing effort to divert scammers and hackers toward synthetic people and computer systems designed to keep criminals engaged while defenders gather intelligence.

One of the most ambitious examples comes from Apate, an Australian company that has spent two years building AI personas for scam calls. The system routes fraudsters to bots trained to sound plausible, show enough uncertainty to invite more persuasion, and avoid actually complying with the scam. The goal is not simply to annoy callers, but to consume time that might otherwise be used to reach real people.

Apate founder and chief executive Dali Kaafar told WIRED that the platform has about 350,000 bots. According to Kaafar, the service is used by banks and supported by telecommunications companies. Its synthetic personas can answer calls, respond to text messages, and enter online scam groups, using different profiles and language abilities so that the interactions do not all follow one obvious pattern.

A protective switchboard sends suspicious calls to many synthetic voice personas.
Apate says its bot network can engage scammers across calls, texts, and online chat groups.

The company also treats those conversations as a source of fraud intelligence. Kaafar said Apate has collected more than 250,000 pieces of information about suspected fraudsters in real time, including scam web addresses, bank details, and accounts used to move money. He said some calls continue for more than two hours, though WIRED did not independently quantify how often that happens across the platform.

WIRED tested a demonstration in which the reporters played scammers pitching a cryptocurrency opportunity to an AI persona. They found the timing natural and the conversation frustratingly credible: the bot remained skeptical without ending the exchange. The test lasted six minutes and did not persuade the synthetic target to invest, an outcome consistent with the product’s purpose rather than proof of its performance at scale.

Scale is the central challenge. Cybercriminals send billions of calls and messages each year, and sophisticated groups operate industrialized scam compounds. Professional scambaiters, researchers, and law-enforcement initiatives have disrupted individual operations for years, but the volume of digital fraud has continued to grow. A bot swarm may absorb some criminal capacity without removing the organizations behind it.

An automated intruder explores an adaptive computer maze while defenders observe from safety.
LLM-powered honeypots can make fake systems more convincing and hold attackers’ attention longer.

A similar strategy is emerging in network defense. Security teams have long used honeypots—decoy machines intended to attract intruders, reveal their methods, and waste their time. Mark Vero, a doctoral researcher in computer science at ETH Zurich, told WIRED that open-source honeypot projects are increasingly adding large language models to make these false systems behave more realistically.

Vero and fellow researchers found that an LLM-powered honeypot kept AI attacking agents occupied significantly longer than a more predictable decoy, according to the report. The attackers were also less likely to identify the adaptive system as a honeypot. Those results suggest generative models can strengthen deception-based defenses, although the article does not establish how the approach performs against every kind of human operator or criminal campaign.

The larger opportunity may be coordination. WIRED notes that police, banks, social platforms, telecommunications companies, and researchers still need better ways to share fraud indicators. AI could help monitor and analyze that information, but collecting clues is useful only if institutions can act on them quickly enough to block accounts, domains, and communication channels.

These defensive systems do not end the asymmetry of online fraud, and people will continue to receive scam calls and messages. Their promise is narrower but practical: turn some criminal automation against itself, make attackers spend scarce attention on convincing fakes, and extract useful signals from the encounter. How much that changes victimization rates remains uncertain, but the economics of deception may no longer favor only the scammers.