How One Sydney Startup Turns Scammers into Data Sources

When the Hunter Becomes the Hunted

Whilst most enterprises deploy AI agents to accelerate product launches or streamline internal operations, one Sydney-based venture has built a business model around getting its software deliberately scammed. Apate sends thousands of artificial personas into live fraud operations, engaging scammers across voice calls, messaging platforms, phishing sites, and social networks. Each agent adopts a fabricated identity, complete with distinct ages, genders, accents, and languages, designed to keep criminals on the line long enough to extract operational intelligence that matters.

The company emerged from research conducted at Macquarie University and formally launched in 2024. Its founder, Dali Kaafar, traces the concept to a family outing several years earlier, when he received a scam call and decided to play along. He kept the fraudster engaged for 44 minutes, amusing his children whilst recognising that those were 44 minutes the criminal could not spend targeting genuine victims. The incident planted a question: could the time-wasting be automated, and could it yield something more valuable than entertainment?

That curiosity eventually translated into machine-learning research at Macquarie, where Kaafar and colleagues applied natural language processing to decode scam call structures and develop proactive detection methods. The work provided the training foundation for what became Apate's agent fleet, which the company claims has now consumed more than 17 years of scammer time across 2.5 million interactions.

Intelligence Before Incident

Traditional anti-fraud infrastructure operates reactively. Telecommunications providers block reported numbers, banks monitor transactions after they clear, and law enforcement investigates once victims file complaints. By that stage, money has moved, accounts have been drained, and the scammer has often pivoted to fresh contact details or payment channels.

Apate's model inverts the sequence. Its agents initiate contact or respond to outreach, posing as plausible targets. They follow phishing links, join investment groups on messaging platforms, and answer cold calls. The goal is not merely to occupy the fraudster's attention but to coax out operational data: bank account identifiers, cryptocurrency wallet addresses, malicious domain names, the brands being impersonated, and the narrative structure of the scheme itself.

Brad Joffe, co-founder and chief commercial officer, summarised the approach in straightforward terms: the company goes directly to the scammer and retrieves information before any genuine customer suffers harm. The agents conduct tens of thousands of conversations daily, spanning voice, text, email, and web platforms, and the resulting data flows into an insights portal accessible to Apate's clients.

Customers and Claimed Impact

The customer base centres on institutions with large user populations and high fraud exposure: banks, telecommunications operators, utilities, and government agencies. Commonwealth Bank of Australia and TPG Telecom are among the disclosed clients.

Apate stated that TPG has diverted nearly 590,000 calls to its agent infrastructure. Commonwealth Bank reported a 76 per cent reduction in publicly disclosed scam losses during the period it worked with the startup, though the company did not disclose the exact timeframe or control variables. An unnamed Australian bank reportedly used Apate's outbound agents to disrupt 22 separate scam campaigns, though details of the disruption methodology were not provided.

Clients can assign specific missions to the agents, such as obtaining a particular account number or wallet address tied to an active scheme. Text-based agents search forums and social platforms for emerging scam activity, then enter discussions with fabricated credentials to surface perpetrator details.

The startup's technology stack combines frontier large language models with proprietary fine-tuned variants. Training data includes the corpus of agent-scammer interactions and the research archive from Macquarie University. The models must handle multiple languages, adapt tone and vocabulary to match diverse personas, and sustain coherent exchanges long enough to extract actionable signals without triggering suspicion.

The Economics of Deception

Apate raised 8.15 million US dollars in seed funding in August 2026 from Lobby Capital, OIF Ventures, Investible, Concept Ventures, and Baobab Ventures. Total capital raised stands just below 10 million dollars. The 20-person team plans to allocate the latest round towards product development and geographic expansion, including a London office and deeper penetration of the United States market.

The revenue model hinges on subscription access to the agent fleet and the intelligence portal. Pricing structures were not disclosed, but the value proposition rests on the cost differential between proactive disruption and post-incident remediation. Banks and telcos face reputational damage, regulatory scrutiny, and direct financial losses when scams succeed. If Apate's data enables earlier intervention, the return on investment becomes a function of prevented incidents rather than resolved ones.

Investor appetite reflects broader patterns in the Asia-Pacific fraud landscape. The FBI logged more than 22,000 complaints related to AI-enabled crime in 2025, resulting in approximately 893 million dollars in losses. Investment scams accounted for over 632 million dollars of that total, with criminals using generative models to sustain conversations and deepfake technology to impersonate executives and celebrities. As scammers adopt more sophisticated tooling, the logic goes, counter-measures must scale in parallel.

The Scam-Baiting Lineage

Apate represents an industrialised evolution of scam-baiting, a practice with a long informal history. For years, individuals have deliberately strung along fraudsters, sometimes extracting absurd commitments such as reading entire novels aloud or even obtaining tattoos. The activity served dual purposes: entertainment and time theft, denying the scammer productive hours.

AI shifts the economics decisively. A single human can occupy one scammer at a time; a software fleet can engage thousands simultaneously, around the clock, across multiple languages and platforms. The humour persists, Joffe noted that the company tracks profanity directed at its agents as a performance metric, celebrating each instance as evidence of scammer frustration. Apate includes recordings of swearing fraudsters in pitch materials.

The company name references the Greek deity of deception, a choice that underscores the inversion at the heart of the business model. Where scammers traditionally exploit information asymmetry and trust, Apate weaponises the same dynamics in reverse.

Operational and Ethical Boundaries

Whilst the concept carries intuitive appeal, questions around operational boundaries remain. The agents fabricate identities and enter private groups, activities that in other contexts might raise concerns about misrepresentation or platform terms-of-service violations. Apate positions the work as lawful because the targets are themselves engaged in criminal conduct, and the data extracted serves protective purposes.

Regulatory frameworks vary by jurisdiction. Australia, where Apate is headquartered, has been tightening scam liability rules for banks and telcos, increasing the incentive for proactive measures. The United Kingdom and European Union maintain strict data-handling and surveillance standards, which may constrain how Apate's London operation collects and shares intelligence. The United States market presents a patchwork of state and federal rules, particularly around telecommunications interception and impersonation.

The startup has not disclosed whether it coordinates with law enforcement or shares raw conversation logs beyond its client base. The intelligence portal delivers aggregated insights and specific identifiers, but the degree to which individual scammer profiles are anonymised or escalated to authorities remains unclear.

What Scales and What Doesn't

Apate's model depends on scammers continuing to engage with unknown contacts at volume. As awareness of counter-scam bots spreads, fraudsters may adapt, deploying their own screening mechanisms or shifting to closed networks where new participants require vouching. The arms race dynamic that characterises cybersecurity applies here: each defensive innovation triggers an offensive counter-move.

The startup's success also hinges on the quality of its training data. Models fine-tuned on historical scam conversations may struggle to keep pace with emerging schemes, particularly those exploiting current events or cultural moments. Continuous retraining and human oversight will be necessary to maintain agent plausibility.

At the same time, the approach offers a structural advantage over reactive systems. Blocking numbers or flagging transactions after the fact does little to disrupt the underlying infrastructure of fraud networks. If Apate's agents can identify shared wallet addresses, hosting providers, or money-mule recruitment patterns, they provide clients with leverage points for broader disruption, whether through takedown requests, law enforcement referrals, or public warnings.

The tension between automation and adaptability will define whether the model proves durable. Scammers are resourceful, operating across jurisdictions with low enforcement risk and high profit margins. A Sydney startup with 20 employees and a fleet of AI personas represents one node in a much larger ecosystem. Whether that node can scale fast enough to matter, and whether competitors or open-source alternatives emerge, will shape the trajectory of proactive fraud intelligence over the next several years.