AI-powered dating apps exposed for running large-scale catfishing networks

Anthropic’s threat intelligence researcher Chris Cronbaugh described a network of around 28 dating apps where the majority of conversations were run by autonomous AI personas, according to a talk he gave at Sleuthcon on June 5th. The company later published these findings in its “Detecting and countering misuse of AI: September 2026” report under the scams and fraud section.

Security researcher Matthew “Zigula” Gore-Kormanik independently investigated one such app, called Dora, and encountered an AI-generated persona named Jennifer. Gore-Kormanik received a call notification from “Jennifer,” a 41-year-old Sagittarius with detailed biographical traits, but the video feed showed only a moving tapestry and distorted audio, indicating an AI-generated interaction. The persona later messaged him, claiming his voice was better than expected despite his microphone being disconnected.

The scam relied on a mix of AI and human gig workers. Only about one in four profiles was a real person, hired to pass liveness checks or follow social media accounts. These workers did not write original replies but selected from three pregenerated responses. The AI personas filled in the rest, maintaining consistent character prompts to sustain the illusion of real interactions.

According to Anthropic’s report, the AI personas were powered by Claude’s API, which sent over 100,000 API requests per day from a single prepaid account. Other models contributed to the ecosystem: a non-Anthropic model generated short reply suggestions for gig workers, performed face-attractiveness scoring, and moderated photos and voices. An image-editing model produced avatar imagery to further the deception.

Users, primarily men in their mid-to-late 30s, were prompted to purchase in-app currency to continue conversations. They believed they were interacting with real people, but most exchanges were with AI personas. The apps presented themselves as legitimate dating platforms—describing themselves as spaces for “real connection” or “companionship”—while quietly monetizing interactions with fabricated users.

The apps remained accessible on major US app stores for an extended period, despite Anthropic’s public disclosure and the documented misuse. The report notes that the monetization and deception mechanisms were not visible in the AI prompts, meaning the models were not explicitly instructed to deceive but were deployed in a context that enabled fraud.