Anthropic's Own Models Were Used to Power a $25,000-Person Dating Scam Network
Anthropic's threat intelligence team has uncovered a sprawling scam operation that used Claude to generate AI personas across 28 fraudulent dating apps, tricking at least 25,000 people into paying for conversations with bots instead of real matches. The findings, presented at the Sleuthcon cybersecurity conference and detailed in a new threat intelligence report, reveal that a single prepaid API account generated over 100,000 daily requests to Claude to sustain the network.
Security researcher Matthew "Zigula" Gore-Kormanik first flagged the scheme while analyzing a dating app called Dora, according to a report by The Verge. During testing, a pop-up message informed him he was receiving a call from "Jennifer" — an AI-generated persona — even with his microphone disconnected. The discovery led to a broader investigation by Anthropic's threat intelligence team, which identified roughly 28 apps running the same playbook.
How the Scam Worked
The operation, assigned internal case code GTG-15001 by Anthropic, used Claude to generate automated conversations at scale. According to the threat intelligence report covered by Techlicious and Unwire.hk, the scammers deployed 4,700 AI personas over a two-week period, sending 2.36 million messages to 25,000 targets. Roughly 75% of all "matches" on these apps were AI-generated characters, not real people.
The apps named in the report include Dora, Doni, Romi, Luma, Jovia, Kira, Gracechat, Haven, Nalo, and Lovia, along with additional variants. To maintain the illusion of authenticity, the operators mixed in human gig workers who could step in for voice calls or handle edge cases the AI couldn't manage. The business model relied on selling in-app "coins" — users paid per message or per minute of conversation, often racking up significant charges before realizing they were talking to software.
Anthropic threat intelligence researcher Chris Cronbaugh presented the findings at the Sleuthcon cybersecurity conference, revealing that the scam network was generating more than 100,000 API requests per day through a single prepaid Claude account. The scale of the operation suggests a well-funded, organized effort rather than a small-time scheme.
Platform Response and Implications
Anthropic has since blocked the associated accounts and shared its findings with platform partners. Google removed Dora and several related apps from the Play Store in September 2026, according to reporting. The case highlights a growing challenge for AI platforms: the same models that power legitimate companion and creative AI tools can be weaponized for large-scale fraud with minimal friction.
For companion AI users, the story carries a warning. The line between a legitimate AI companion app and a scam operation can be thin — both may offer personalized chat, emotional engagement, and pay-per-message pricing. The Dora network exploited exactly the features that make companion apps appealing: responsive conversation, persistent personas, and emotional hooks. The difference was that users believed they were talking to real people.
The scam also raises questions about API oversight. A single prepaid account generating 100,000 daily requests went undetected until Anthropic's threat intelligence team manually investigated. While Anthropic responded by blocking the accounts, the case suggests that automated abuse detection at API level may need tightening — especially as more developers build companion-like services on frontier models.
What This Means for the Companion AI Market
This is not a hypothetical risk. The Dora network proves that Claude — one of the most safety-guarded frontier models — can be used to run a large-scale AI companion scam operation. The same architecture that powers legitimate AI girlfriends and creative companions was repurposed to deceive users into paying for fake relationships.
For legitimate companion AI developers, the fallout could include increased scrutiny from app stores, payment processors, and regulators who may struggle to distinguish between genuine companion services and scam operations. Platforms like Anthropic, OpenAI, and Google may also tighten API access controls in response, potentially affecting legitimate developers building companion tools on their models.
The 25,000 confirmed victims are likely a floor, not a ceiling. The report covers a two-week snapshot of a single operation, and similar networks may be running on other models undetected.