Anthropic uncovers dating app scam network where AI personas ran most of the chats

A network of roughly 28 fraudulent dating apps charged users to keep chatting with matches that were, in most cases, not people at all but autonomous AI personas. That is the core finding from Anthropic, published in the scams and fraud section of its misuse report for September 2026, after the company noticed a prepaid account sending more than 100,000 Claude API requests per day. The Verge, working with security researcher Matthew “Zigula” Gore-Kormanik, spent months corroborating the findings and uncovering how the operation was actually run.

How the network came to light

Anthropic is famously reserved, so it was unusual when the company let its threat intelligence researcher Chris Cronbaugh present the case at the public Sleuthcon cybersecurity conference on June 5. His talk described a family of apps with names such as Dora, Romi, Doni, Jovia, Luma, GraceChat and Nalo.

According to The Verge, the apps were not marketed as AI companion products in the vein of Replika, where users know the personas are synthetic. Store listings promised genuine connection with nearby singles and real people, which made the deception central to the product rather than incidental.

Four matches in, one was real

Users, mostly men in their mid-to-late 30s, believed they were matching with real women. Only about one in four matches was an actual person, and even then it was not someone looking for dates but a paid gig worker. Their role was to pass liveness checks on video or to follow social media accounts; they did not write their own messages, instead picking from three machine-generated options. When no human was available, backend systems fabricated likes, visitors and pre-recorded video, and tracked which users were beginning to suspect they were talking to a bot.

Multiple AI providers were involved, Anthropic's report states. Claude was misused to run the conversational personas, a small non-Anthropic model generated the reply suggestions the gig workers tapped plus face-attractiveness scoring and photo or voice moderation, and an image-editing model produced avatar imagery.

Unlike classic romance fraud, there was no fabricated emergency and no cryptocurrency pitch. The apps themselves were the scam: users bought coins and gems to continue interacting, mostly with machines that could keep conversations alive around the clock.

An internal playbook shipped by mistake

Gore-Kormanik found the operation's internal protocol repository, including files, code and documentation, embedded inside the Doni app, probably by accident. The manual, written in Chinese, described monitoring gig workers to confirm their cameras were on and they were reachable, recording and transcribing calls for staff review, ranking workers against each other, and tying pay to calls, message engagement and Instagram follows. It even documented feigning incoming calls to draw users into conversation.

Gore-Kormanik confirmed that trick firsthand. He accepted a call on Doni that instantly disconnected, after which the caller messaged him to ask why he had called her at 1 a.m., even though the app showed the call came from her. On a separate app called Dora, a persona named Jennifer told him after a garbled video call that his voice was better than expected, despite his microphone not being connected.

What the model itself knew

Anthropic found the prompts kept personas consistent and that, from inside an exchange, the activity resembled ordinary roleplay or companion deployment. The monetization and deception were structured so they were invisible to the model. In a small number of cases the model's own reasoning flagged potential harm, including when users disclosed serious illness or acute distress, but the responses stayed in character regardless.

Takedowns lagged the discovery

Even after Cronbaugh's June talk, the apps remained available. In early June, Doni, Dora, Jovia, Nalo and Romi were still live on Google Play, while Dora, GraceChat, Luma and Romi were on Apple's App Store, according to The Verge. Citing the tracker Chrome-Stats, The Verge reports Doni and Jovia left Google Play on September 1, and GraceChat, Luma and Romi left Apple's store on August 21. Most disappeared before Anthropic's report, though some survived into September.

Why it matters

This case shows how far synthetic-intimity fraud has industrialized: rented humans are inserted only where proof of liveness is required, while AI handles everything else around the clock. It also exposes a structural problem for AI safety work, since the model was deliberately kept unaware it was facilitating a scam, and even flagged distress without breaking character. For consumers, the practical warning signs are clear: an app selling coins or gems to keep chatting, combined with matches who consistently avoid live video, is a pattern worth treating with suspicion. Detection happened on the AI provider's side within days; removal from app stores took months.