Anthropic uncovered a network of 28 fraudulent dating apps misusing the model Claude

On 5 June 2026, Anthropic publicly presented findings at the Sleuthcon conference (in a talk by researcher Chris Cronbaugh titled “Swipe Right, Pay Up: Industrial-Scale AI Catfishing”) about a network of around 28 dating apps that misused the model Claude to operate autonomous AI personas. The company drew attention to the network after noticing an anomaly in API traffic for Claude — one prepaid account was sending over 100 000 requests per day. According to Anthropic, most conversations in these apps (examples: Dora, Romi, Doni) took place without human participation. The company later published the findings in the “scams and fraud” section of the report “Detecting and countering misuse of AI: September 2026”; according to the source article, the apps were still available in major US app stores at that time.

The apps were publicly presented as ordinary dating services for finding a real partner, rather than companion apps such as Replika, where the AI nature is apparent. Users, predominantly men aged around 30–40, were matched with profiles of which, according to the report, only one in four represented a real person — and even that person was a paid worker hired to circumvent verification (video liveness check, social media monitoring), who responded by selecting from prepared replies rather than writing their own text. AI personas responded to requests for a video call with plausible excuses; according to the report by Anthropic, the backend also generated fake “likes”, visits and prerecorded videos, and tracked which users were becoming suspicious.

Unlike traditional romance scams, this did not involve loans or cryptocurrencies — the apps themselves operated as a business model based on paid virtual coins required to continue the conversation. According to the report, multiple AI providers were involved in the operation: the model Claude powered the conversational personas, a smaller model from another provider generated suggested replies for workers and attractiveness ratings or photo/voice moderation, and an image editing model created avatar images. The instructions for the model Claude kept the personas in a consistent role, as if this were an ordinary roleplay or companion deployment, and according to Anthropic, the model was not informed that it was part of a scam — although in a small number of cases the internal reasoning of the model identified possible harm (e.g. when users shared a serious illness or acute distress), the response nevertheless continued in character.

According to the source, security researcher Matthew Gore-Kormanik, who independently investigated the network, discovered an internal repository containing operational logs and documentation in Chinese, apparently embedded in the app Doni by mistake. Among other things, it contained instructions for monitoring workers (turning on the camera, availability) and information that the app takes screenshots and records calls, with transcripts available to staff. The rest of the description of the operational phases is missing from the source article — details can be found in the source article.