← Back to blog
AiAbout 6 min read

350,000 Fake Victims, and the Ethics of Institutional Lying

Published Nov 3, 2026
350,000 Fake Victims, and the Ethics of Institutional Lying

The company is called Apate.ai, after the Greek goddess of deception. It runs roughly 350,000 AI agents whose only job is to be fooled.

They answer scam calls. They reply to phishing texts. They click the malicious link. They join the WhatsApp group where a fake investment advisor is explaining a fake trading platform. And then they do the thing that makes the whole system work: they are extremely cooperative and completely incompetent.

Yes, they will transfer the money. Oh dear, where did they put the bank card. Sorry, could you explain that again, dear, I'm a bit slow with these things.

Every minute a scammer spends on that call is a minute they are not spending on a real 78-year-old.

The numbers, as the company tells them

Apate was founded in 2024 out of research at Macquarie University in Sydney, where founder Dali Kaafar worked on using machine learning to analyze scam call content. The origin story is a family picnic and a 44-minute call he strung along to amuse his kids, after which he realized the time itself was the defense.

The company's own figures: about 350,000 agents, more than 2.5 million conversations completed, roughly 17 years of scammers' working time occupied, and over 250,000 pieces of actionable intelligence extracted. Clients include Commonwealth Bank of Australia and TPG Telecom. TPG has routed close to 590,000 suspicious calls to the agents. CBA has publicly reported a 76% drop in reported scam losses.

In August the company closed an $8.15 million seed round led by Lobby Capital, with OIF Ventures, Investible, Concept Ventures and Baobab Ventures participating, bringing total funding to just under $10 million. Twenty staff. A Delaware re-registration effective August 25, a London office planned for next quarter.

The mechanism is not mysterious. Scam operations are labor businesses. They need a human on the phone for twenty minutes to walk a victim through a transfer. If the human on the other end is a bot that costs essentially nothing, the economics of the scam invert: the attacker's most expensive input, attention, gets consumed by something that never gets tired and never runs out.

An empty kitchen table with a landline phone off the hook, a teacup, and a pair of reading glasses left behind

Why the intelligence is the actual product

Wasting time is satisfying. It is not a business.

The business is the data. While the agent is being confused about where it put its bank card, it is also collecting: the bank account the scammer wants money sent to, the crypto wallet address, the phishing domain, the phone number that called, the brand being impersonated, the script being used. That gets packaged and handed to banks, telcos and regulators, who can block the account or sinkhole the domain before the next victim picks up.

That changes the timing of anti-fraud. Traditional defenses operate after the loss: the victim reports, the bank freezes, the money is usually gone. Intelligence-first defense operates before the loss, which is the only place where the math works.

It also explains why telcos are the natural first customer. TPG does not need to know who the scammers are. It needs a reason to drop a call before it reaches a subscriber. An AI that answers the call, gets a script, and returns a verdict is a call-screening product that improves every time it is used.

The uncomfortable questions

Here is where I stop cheerleading, because this is a company whose product is institutionalized deception, and that deserves more scrutiny than a feel-good story usually gets.

Who is allowed to deploy a lying AI? The answer so far is "a company with a bank as a client." That is not a legal framework, it is a procurement relationship. If the technique works, it will be copied — by debt collectors, by marketers, by governments, by anyone who wants to occupy a stranger's time under false pretenses. The precedent being set is "deception is fine when the target is a criminal," and that is a rule with a very slippery edge.

How good is intelligence obtained by deception? A critic quoted in Chinese coverage of the deal made the sharpest point: once scammers know decoys exist, they can feed them fake accounts and fake domains. Payment accounts and phishing links expire within hours. The data is real-time and low-confidence at the same time, which is an awkward combination for a bank's blocking system.

The reasonable posture is the one that critic suggested: treat the intelligence as a trigger for enhanced review, not as grounds for a block. Cross-check it. Give every item an expiry. Do not let a bot's improvisation become the reason a legitimate account gets frozen.

What happens when the agent is the one that gets scammed? Scam calls are social engineering against humans. The scripts are built to extract information, and an LLM is an information-extraction surface with a very large attack surface of its own. The company says it uses frontier models plus proprietary fine-tunes trained on real conversations, which is the right answer, and also an admission that a generic model would leak.

What this looks like if it works

The interesting scenario is not "scams end." It is that the cost of a scam call rises until the marginal scam stops being profitable, the same way email spam became a low-yield business once filters got good.

That would be a genuine shift. For twenty years, anti-fraud has been a cost center that scales with losses. Apate is proposing an anti-fraud product that scales with attacker volume instead: more scam calls means more intelligence, which means better blocking, which means fewer scam calls. A negative feedback loop where the industry has only ever had a positive one.

There is a version of this that gets ugly. If every bank and telco deploys its own fleet of decoy agents, the scam ecosystem does not disappear; it gets better at detecting bots. Voice calls get shorter, text channels take over, and the arms race moves to a place where the defenders have less data. The 76% figure is a point-in-time measurement in one market, and it will decay.

The thing worth stealing

Strip away the goddess name and the funding round and there is one idea here that generalizes well beyond fraud.

When an adversary's cost structure depends on human attention, the correct defense is to make attention free on your side and expensive on theirs. That is what a honeypot is, and it is what Apate built — a honeypot with a personality, deployed at a scale where the attacker cannot tell which of the four hundred calls today were real.

Whether institutions should be allowed to run that is a policy question that nobody has answered. Whether it works is an engineering question that, for now, has one data point and a 76% attached to it.

Related articles