200,000 fake AI victims trap online scammers

AI scams have evolved from novel tool to infrastructure

Australia company Apate claims that it has deployed nearly 200,000 AI "characters" around the world, imitating gullible targets-both stalling fraudsters and gathering intelligence for banks and telecom operators. In a six-week period ending in 2025, Apate reported that its robots participated in 600,000 fraudulent calls by Australia telecommunications company TPG, which equates to wasting more than 500 days of criminals 'time-saving approximately US$13 million. In addition to disruptive effects, Apate's system is designed to extract actionable details to help defenders identify where stolen funds are going and how fraud networks operate.

Key Points

Apate said it operates nearly 200,000 AI characters and can interact with fraudsters through phone-based conversations and messaging channels.

In the six weeks to the end of 2025, Apate reported processing 600,000 fraudulent calls for TPG, which is equivalent to wasting hundreds of days of fraudsters 'time.

The company's anti-fraud value proposition is not just about procrastination, but also focuses on extracting intelligence, such as new cryptocurrency wallet addresses.

Apate's research shows that AI is increasingly involved in fraudulent communications, but it believes defenders can still have an advantage.

This approach reflects broader industry efforts, including British Telecom's O2 's "AI grandma" campaign, but Apate uses its size and intelligence extraction capabilities as a differentiator.

AI targets designed to waste fraudsters 'time

Apate founder Dali Kaafar described the system as an automated way to achieve the long-term goal of human scams 'manual operation: by deceiving scammers to consume their energy, rather than the public's. According to Kaafar, the idea originated from a fraudulent phone call he received during a family picnic in Sydney in November 2021. He spent 44 minutes dealing with the caller by playing an innocent victim.

What started as a private pastime evolved into a research-led project. While serving as a professor at Macquarie University, Kaafar discussed the concept with doctoral students focusing on AI and security, proposing a system that can both interact with fraudsters on a large scale and capture useful information from those interactions.

Within months, the project received research funding from the Office of National Intelligence and was subsequently spun off to form Apate in 2023. Kaafar said the company currently works with major banks in Australia and other financial institutions in the UK, South Africa and parts of Southeast Asia.

Nearly 200,000 characters and "realistic" dialogue behavior

Apate started with 120 different characters and later expanded to 197,000. Kaafar attributes the realism to detailed character setting, including recognizable tone characteristics, accents and subtle behavioral clues. He said the company spent a lot of time optimizing the robot's voice and response so that the conversation felt natural to the target-and convincing enough for a potentially skeptical scammer.

According to Kaafar, the underlying AI model was trained on recorded conversations between "hundreds" of human scams and scammers. This training is aimed at implementing adversarial strategies during phone calls and chats, rather than just running automated scripts.

Apate also uses the same interactive cycle as a measurement tool. Kaafar said an internal performance indicator tracks how often frustrated scammers use swear words on robots-a detail that highlights the company is optimizing ongoing interactions rather than hanging up quickly.

The company deploys robots through channels such as WhatsApp and Telegram, where fraudsters often try to quickly move from initial contact to payment instructions. Kaafar also outlines the strategy around a key behavioral truth: Even if "you can't fool honest people" is not literally true, scammers are still driven by greed, and this motivation can still be exploited by delaying or steering their work processes.

From interference to defense: extracting cryptocurrency and operational intelligence

Apate describes its anti-fraud data goal as providing forward-looking intelligence to banks and telecom operators. In cryptocurrency-related work, Kaafar said Apate works with "one of the leaders in blockchain analytics" and is committed to identifying wallet addresses and methods used by fraud groups.

Kaafar claims that in multiple conversations with robot interactions, Apate can extract new cryptocurrency wallet addresses "hundreds of thousands". The underlying logic is that defenders need to identify where funds will fall next before they are transferred-thereby applying monitoring and incident response before damage occurs.

He likened fraud operations to corporate organizations, implying that call centers and related workflows are sufficiently structured to support repeatable data advantages. In this framework, the most important output of robot interaction is not just evidence that fraud has occurred, but also crucial details that help analysts understand which accounts, wallets, and compound "payment points" are next.

For example, Apate said that in July its robots discovered a market involving brokers soliciting India to verify bank accounts, who reportedly received commissions paid in USDT based on fraud proceeds flowing through those accounts.

An arms race in which defenders may still have the upper hand

Apate's broad warning is that fraudsters are also increasingly using AI. The company pointed out that because fraud itself is already a huge business, fraud can be scaled at low cost. Kaafar said Apate's research estimates that approximately 20% to 30% of fraudulent text messaging conversations already use AI, which reflects how quickly fraudsters can adopt tools to speed up communications.

Despite this, Kaafar believes that anti-fraud robotic systems have strategic advantages in the AI confrontation between defenders and attackers. Drawing on game theory concepts, he proposed that defensive robots are designed to extract information, while fraud robots try to push the other party into action-meaning that defenders can more easily learn from the attacker's models and behaviors.

Kaafar also takes a longer-term view: Even as fraudsters become more complex, this complexity may increase the amount of usable data left behind in interactions. He described the development as "good news" in the fight against fraud.

The key is not that AI eliminates fraud risk, but that well-designed interactive systems can turn fraud attempts into intelligence streams-converting wasted victim time into resources for investigators and surveillance teams. [TAG

As AI scams evolve, readers should be concerned about whether robot-based intelligence extraction will become standard for financial institutions and telecom providers-and, more importantly, whether defenders can quickly act on wallet and account details discovered by these systems before transfers occur.