It started with a single, suspicious SMS sent to a UK phone number that happened to belong to an OpenAI investigator. That lucky break allowed OpenAI to map out and shut down a centralized scam operation they’ve dubbed “Wrong Number,” which was using ChatGPT to run a cross-platform task scam likely originating from Cambodia. The actors were banned for generating recruitment spam and translating conversational scripts across six languages, including English, Spanish, Swahili, and Haitian Creole.
The model’s role was functional and strangely mundane. The network didn’t use ChatGPT to craft hyper-personalized deepfakes but to perform bulk translation: converting incoming messages into Chinese for operators and then pushing replies back out in the target’s native language. This allowed a centralized, Chinese-speaking team to scale their con globally without speaking the victim’s language. The actual cold-contact SMS and initial WhatsApp messages were boilerplate and, notably, not AI-generated—the real grunt work for the scam didn’t need a large language model.
The scam itself followed a depressingly classic “pig-butchering” adjacent pattern OpenAI breaks down as ping, zing, and sting. The ping was an out-of-the-blue job offer, often via SMS, promising absurdly high wages for liking social media posts—one message offered over $5 for a single TikTok like, a rate wildly out of step with the real market where 1,000 likes go for under $10. The zing involved translating pep talks and bonus announcements to keep new recruits engaged. The sting was the endgame: pressuring victims to pay deposits, crypto fees, or “handling charges” to unlock the promised payouts. Fake companies like Hyesung Advertising and Lightning Shared Scooter Co. were repeatedly used as fronts.
While OpenAI can’t quantify the full body count, the evidence of real-world harm is clear. Their investigation found conversations where “employees” demanded refunds and social media posts from genuine users defending the fraudulent companies. This hybrid operation—human-run coordination powered by AI for linguistic scale—is a far cry from fully autonomous AI threats. It’s a more immediate and boring reality: ChatGPT used as a cheap, on-demand translation layer for an otherwise traditional scam, proving that the most effective malicious use of AI right now isn’t about sophistication, it’s about industrialization.
