Turning the Tables on Cybercrime: How AI "Honeypots" and Bot Swarms Are Fighting Back Against Digital Scammers

By Lily Hay Newman and Matt Burgess
Special to Kernel Panic

Main Facts: The AI Paradox in Modern Cybersecurity

Artificial intelligence has undeniably supercharged the landscape of cybercrime. Today, bad actors leverage generative tools to craft hyper-realistic phishing emails, deploy automated malware, scale up social engineering campaigns, and generate deepfake audio and video to trick unsuspecting victims out of their life savings. The global crisis of online fraud has grown exponentially, leaving international law enforcement agencies struggling to keep pace.

Yet, in a classic case of fighting fire with fire, this very same technology is now being recruited on the front lines of defense. Governments, telecommunications giants, financial institutions, and academic researchers are turning the tables on fraudsters. Instead of merely playing defense—blocking numbers or updating spam filters—cybersecurity innovators are deploying sophisticated artificial intelligence to actively waste scammers’ time, drain their operational resources, and harvest real-time intelligence.

At the heart of this counter-offensive are AI-driven "honeypots" and conversational bot swarms. By simulating the perfect, vulnerable targets, these systems engage fraudsters in multi-hour phone calls, text message threads, and online chat groups. The primary objective is simple yet devastatingly effective: every minute a scammer spends talking to an artificial intelligence bot is a minute they cannot spend defrauding a real human being.

Chronology: The Evolution of Automated Counter-Offensives

The war against online fraud has evolved through several distinct phases over the past two decades, leading to the current era of generative AI defense systems.

• Early 2000s: The Rise of Traditional Honeypots. Security researchers and corporations began deploying static virtual machines and decoy systems to attract hackers. These early honeypots were designed to log unauthorized access, observe hacker behaviors, and alert defenders to emerging malware strains. However, they were easily identifiable by sophisticated attackers due to their rigid, predictable responses.
• The "Scambaiting" Era (2010s). Independent digital vigilantes—known as scambaiters—began independently wasting fraudsters’ time by pretending to fall for advance-fee scams. While effective at a micro-scale and heavily entertaining on platforms like YouTube, these efforts were entirely manual, limited in scope, and could not scale to combat industrial-sized scam compounds.
• 2022: The Birth of Apate. Named after the Greek goddess of deception, the Australian anti-fraud company Apate was founded to industrialize the concept of wasting scammers’ time. Over the last two years, the firm built an automated telephony and text-based system designed to divert phone scammers onto calls with hyper-realistic AI bots.
• 2023–Present: Generative AI Integration. Academic and industry researchers began integrating Large Language Models (LLMs) into honeypots and defensive systems. Recent academic breakthroughs, such as research conducted at ETH Zurich, proved that LLM-powered deception frameworks could trick automated and agentic cybercriminal tools far more effectively than traditional scripts, forcing a paradigm shift in how defenders interact with attackers.

Supporting Data: The Scale of the Bot Counter-Offensive

The numbers behind these emerging AI-driven defense mechanisms reveal an industrial-scale effort to match the operational capacity of cybercrime syndicates:

• 350,000: The approximate number of active AI bots deployed by Apate, working in tandem with major banks and telecommunications providers to intercept fraudulent communications.
• 250,000+: The volume of real-time intelligence data points harvested by Apate’s bot swarms. This includes malicious URLs, operational phone numbers, cryptocurrency wallet addresses, and illicit money mule bank accounts.
• 2+ Hours: The average duration of individual scam calls handled by Apate’s AI agents, proving that the personas are convincing enough to keep fraudsters on the hook for extended periods.
• Zero: The number of successful cryptocurrency investments secured during Kernel Panic’s rigorous live testing of Apate’s demo system, despite journalists tag-teaming the AI as a duo consisting of a friend named "Lucy" and her financial advisor "Mickey."

Furthermore, academic studies evaluating LLM-integrated honeypots—such as the recent work by Mark Vero and his colleagues in the Department of Computer Science at ETH Zurich—show that automated, agentic cybercriminal tools remain engaged with LLM-simulated honeypots "significantly longer" than traditional, predictable honeypots, identifying them as genuine targets at a much lower rate.

Official Responses and Insights from the Front Lines

The deployment of conversational AI against cybercriminals represents a fundamental shift in mindset: treating fraudsters not just as code to be blocked, but as human actors with psychological vulnerabilities.

Dali Kaafar, the founder and CEO of Apate, describes the company’s core philosophy with a touch of irony:

"What we really like to think is that we’re building the perfect victims for scammers. A minute that a scammer is talking to a bot or an agent is a minute where you’re probably saving hundreds, if not thousands of possible people being reached out to by that exact same scammer."

Kaafar emphasizes that Apate’s platform is engineered to mimic human unpredictability. The bots are provisioned with diverse personalities, varying linguistic capabilities, and realistic behavioral traits. Sometimes a bot will use WhatsApp; sometimes it won’t. Sometimes it answers a call immediately, and other times it drops the call with a casual excuse—"I’ll come back to you later"—mirroring the imperfect habits of real mobile phone users.

Testing the system firsthand reveals the psychological nuance built into these models. Journalists from Kernel Panic tested a demo version of Apate’s AI "victim" personas. They found that the system strikes a delicate balance: it expresses a healthy, natural level of skepticism while leaving just enough conversational openings to encourage the scammer to keep trying.

On the technical research side, Mark Vero of ETH Zurich underscores the strategic advantage this gives to system defenders. Reflecting on recent honey-pot evaluations, Vero noted:

"The agentic attackers are much more convinced by the LLM-simulated honeypots, and they also mark them as actual honeypots at a much lower rate. If these systems are built well enough, then I think it’s quite advantageous for defenders."

Implications: The Future of Anti-Scam Operations

Despite the impressive scale of Apate’s bot swarm and the technical sophistication of LLM-powered honeypots, experts acknowledge that these tools are currently operating as a thumb in the dike against a rising global tide. Cybercriminals initiate billions of malicious messages and calls annually, often operating out of massive, industrial-scale scam compounds spanning multiple international jurisdictions.

Traditional law enforcement and isolated cybersecurity initiatives have historically struggled to dismantle these networks due to jurisdictional boundaries and the sheer velocity of digital crime. However, the integration of generative AI into defense paradigms opens up several profound implications for the future:

1. Exploiting Psychological Vulnerabilities: As security researchers increasingly lean into "psyops" against cybercriminals—ranging from automated time-wasting bots to directly trolling ransomware operators—the psychological cost of running scams increases. If fraudsters spend half their day talking to AI bots that yield zero financial return, the economic model of cybercrime begins to fracture.
2. Enhanced Intelligence Sharing: The rich data harvested by conversational bots—such as money mule accounts and scam infrastructure URLs—provides critical intelligence. If effectively funneled to financial institutions, social media platforms, and law enforcement agencies, this data can be used to preemptively freeze illicit funds and take down infrastructure before mass victim impact occurs.
3. A New Paradigm in Automated Defense: The success of LLM-backed honeypots signals a broader industry transition. As attackers increasingly deploy autonomous AI agents to scan and exploit networks, defenders must respond in kind with autonomous deceptive agents designed to confuse, exhaust, and outmaneuver them.

While it remains only a matter of time before your phone buzzes with yet another fraudulent text or robocall, the tools being forged today suggest that the future of digital security will not be entirely defensive. By weaponizing artificial intelligence to waste the time and exploit the psychology of cybercriminals, defenders are finally giving scammers a taste of their own medicine.