Daniel Spokoyny has spent the better part of two years teaching machines to waste fraudsters' time.
It's a peculiar pursuit—one that involves dispatching AI chatbots to flirt with romance scammers, feign interest in dubious crypto investments, and string along con artists long enough to extract the digital breadcrumbs they leave behind. Mule bank accounts. Fake trading platforms. Cryptocurrency wallets. The entire apparatus of modern fraud, one tedious conversation at a time.
"We're essentially running honeypots at scale," says Spokoyny, a Carnegie Mellon-trained machine learning researcher who co-founded BeeSafe AI last year. The San Francisco startup, which emerged from Y Combinator's Winter 2026 batch, is betting that the best defense against trust-based scams isn't better anomaly detection—it's turning the tables entirely.
The timing feels urgent. Americans lost $12.5 billion to fraud in 2024, according to Federal Trade Commission data released earlier this year. Investment scams alone accounted for $5.7 billion. And those figures probably understate the problem, given how many victims never report losses out of embarrassment or resignation.
Traditional fraud prevention tools stumble here because the transactions look legitimate. Victims authorize the payments themselves, often after weeks of carefully orchestrated manipulation on WhatsApp or Telegram. By the time a bank's system flags anything unusual, the money has already landed in a mule account halfway across the world, ready to vanish into a labyrinth of crypto mixers and offshore exchanges.
BeeSafe's approach is different—arguably stranger. Rather than wait for victims to come forward, the company deploys what it calls "Anti-Scam Agents" to engage fraudsters directly, posing as marks to extract intelligence on the financial plumbing behind pig butchering schemes, romance scams, and impersonation fraud.
When the Bait Bites Back
The concept isn't entirely new. Scambaiting—the practice of deliberately wasting scammers' time—has existed in various forms for years, mostly as vigilante performance art on YouTube or Reddit. What BeeSafe claims to have built is something more systematic: an AI-powered intelligence engine that can hold coherent, trust-building conversations across thousands of simultaneous engagements, then parse the results for actionable data.
The three-person founding team brings serious academic firepower. Spokoyny holds a PhD in machine learning and natural language processing from Carnegie Mellon and serves as principal investigator on a $305,000 National Science Foundation grant focused on "Counteracting Social Engineering Attacks with Honeypot LLM Chatbots." His co-founders are Ariana Mirian, who completed her security PhD at UC San Diego in 2023 after stints as a senior security researcher at Censys, and Nikolai Vogler, another UCSD PhD with roots in CMU's Language Technologies Institute.
Their academic work provides the foundation. Last October, the team published a preprint titled "Victim as a Service: Designing a System for Engaging with Interactive Scammers," which detailed a system called CHATTERBOX. A follow-up study from September 2025 reported that LLM-based scambaiting achieved a 32% information disclosure rate in large-scale tests—meaning roughly one in three scammers coughed up useful intelligence during engagements. The human acceptance rate hovered around 70%, suggesting most fraudsters couldn't distinguish the bots from real targets.
Those numbers sound impressive until you consider the operational headaches. Engagement bottlenecks. Variable disclosure rates depending on scam type. The challenge of maintaining evidence chains that would hold up in court or satisfy compliance officers. The research flagged all of it.
BeeSafe claims to have logged tens of thousands of real scammer conversations and identified thousands of mule accounts and linked infrastructure. That intelligence, the company says, flows to financial institutions, cryptocurrency platforms, telecom providers, and government agencies. Banks use it to detect money mule accounts before authorized push payment fraud drains customer accounts. Telcos block suspicious inbound traffic. Law enforcement gets leads for takedowns.
Whether those claims hold up at scale remains an open question.
The Economics of Trust-Based Fraud
The problem BeeSafe is tackling has metastasized over the past few years, fueled by pandemic isolation, the mainstreaming of cryptocurrency, and increasingly sophisticated social engineering tactics.
Pig butchering—a term borrowed from Chinese fraud parlance—epitomizes the trend. Scammers spend weeks or months cultivating romantic relationships with targets, often on dating apps, before gradually introducing investment opportunities. The schemes lean heavily on fake cryptocurrency trading platforms that show fabricated returns, encouraging victims to deposit more funds. When victims try to withdraw, the platform vanishes.
The FBI's Internet Crime Complaint Center logged internet crime losses exceeding $16 billion in 2024, up 33% year-over-year. Investment and crypto scams drove outsized damage. Romance and pig butchering scams grew roughly 40% last year and represented more than a third of crypto scam revenue, according to Chainalysis data cited by the Independent Community Bankers of America. Total crypto scam losses hit $9.9 billion.
These attacks bypass traditional fraud defenses because they exploit human psychology, not system vulnerabilities. Victims authorize the transactions. They believe they're investing, helping a romantic partner, or claiming a prize. Anomaly detection flags unusual patterns—but what's unusual about a customer sending money to an investment account they opened themselves?
Regulatory responses are starting to shift the burden. The UK's Payment Systems Regulator implemented mandatory reimbursement for authorized push payment fraud in October 2024, requiring banks to reimburse victims within five business days, with a cap of £85,000. Early metrics from the first quarter showed high reimbursement rates, creating strong incentives for both sending and receiving institutions to prevent fraudulent transfers before they happen.
In the U.S., FinCEN issued an alert in September 2023 detailing pig butchering red flags and encouraging Suspicious Activity Reports. The guidance underscores how intelligence enumerating mule accounts and scam infrastructure aligns with existing anti-money laundering and transaction monitoring workflows.
Translation: banks are suddenly very interested in any tool that can help them avoid eating losses.
A Crowded Field
BeeSafe is hardly alone in chasing this opportunity.
The mule detection market already includes behavioral biometrics provider BioCatch, which launched its Mule Account Detection solution in 2021, along with established fraud prevention players like Featurespace, Feedzai, NICE Actimize, and FICO. Cybera offers an "AI Scam Engagement System" and "Mule Intelligence" product with positioning similar to BeeSafe's—verified, non-probabilistic intelligence gathered through direct interactions.
BeeSafe's pitch centers on what it calls "full campaign context" and "evidence-backed, verified intel" that minimizes false positives. The company emphasizes real-time discovery and cross-channel mapping that links initial contact through to fund exfiltration, delivering signals to customers without adding friction for end users.
That last point matters. Fraud prevention teams live in constant tension between security and user experience. Flag too many legitimate transactions, and you alienate customers. Miss too many fraudulent ones, and you're liable for the losses.
The company is backed by Y Combinator and lists Obvious Ventures and America's Seed Fund SBIR/STTR among its supporters. Obvious recently closed a $360 million Fund V in January. BeeSafe has not disclosed customer names, detailed pricing, or revenue figures—standard opacity for an early-stage startup still building out its go-to-market motion.
The Scalability Question
BeeSafe's platform is currently accepting requests for early access through its website, which was updated February 12. The company claims to have already enabled financial services firms and government agencies to intercept scammers before transactions complete, though it hasn't named these organizations publicly.
The real test will be scalability—not just technical, but operational.
Academic research on automated scambaiting systems points to persistent challenges: engagement bottlenecks when scammers demand video calls or other proof of identity, variable disclosure rates across different scam types, and the need for human oversight to maintain evidence integrity. Those challenges matter when you're trying to move from tens of thousands of conversations to the millions required to make a meaningful dent in a $12.5 billion problem.
There's also the cat-and-mouse dynamic. Scammers adapt. Once they realize AI agents are wasting their time, they'll develop countermeasures—verification tests, subtle conversational tripwires, patterns that distinguish bots from real marks. It's the same arms race that plays out across cybersecurity, just in a different theater.
For fraud executives facing mandatory reimbursement regimes and rising losses, the value proposition is straightforward enough: verified intelligence on mule accounts and scam infrastructure, extracted before victims lose money. Whether AI agents can deliver that intelligence at the volume and accuracy required to shift the economics of trust-based fraud—well, that's the experiment BeeSafe is running.
In the meantime, somewhere on WhatsApp, a chatbot is probably flirting with a romance scammer, playing along just long enough to learn which bank account comes next.
