How an OpenAI Investigator's Own Phone Helped Bust an AI Task Scam

An SMS offering easy money landed on the phone of an OpenAI investigator — and that accidental delivery gave the company a direct look into a ChatGPT-powered task scam network it now calls Operation "Wrong Number."

The message, generated with ChatGPT, offered high hourly pay for trivial work such as liking social media posts. OpenAI disclosed the operation in its June 2025 report on disrupting malicious uses of AI, where it published the case study as part of a broader accounting of banned accounts and disrupted campaigns.

OpenAI banned the ChatGPT accounts behind the scheme. According to the company, the accounts produced short recruitment-style messages in English, Spanish, Swahili, Kinyarwanda, German, and Haitian Creole. The messages offered recipients high salaries for minimal tasks and pushed them to recruit others. "The operation appeared highly centralized and likely originated from Cambodia," OpenAI wrote.

The case matters beyond one busted network. Task scams — in which victims are paid small amounts at first, then pressured to deposit money to "unlock" larger rewards — have grown into a documented consumer threat, with the FTC warning about them as recently as November 2024. Operation "Wrong Number" shows how commercial AI tools can industrialize the multilingual outreach and translation work that such frauds demand, while also showing how the same tools can expose the operation to detection.

A translation pipeline for fraud

Most of the banned network's ChatGPT activity involved translation. The operators used the model to move conversations between Chinese and six other languages: English, Spanish, Kinyarwanda, Swahili, German, and Haitian Creole. The tasks alternated — translating an incoming message into Chinese, then translating the operator's Chinese response back into the recipient's language.

OpenAI said its own AI-powered translation tools helped it investigate and disrupt the campaign quickly. Scammers and investigators were, in effect, using the same class of technology against each other.

The recruitment funnel ran across multiple messaging platforms. The initial SMS directed recipients to WhatsApp, where responders were routed to a "mentor" on Telegram. Some activity also referenced the BonChat messaging app. OpenAI noted that the early introductory messages on these channels did not appear to be model-generated — they consisted of routine text that would likely apply to every conversation.

Fake employers, familiar names

Many translated communications purported to come from representatives of "employer" companies. The same names surfaced repeatedly across ChatGPT conversations in the investigation: Hyesung Advertising and Lightning Shared Scooter Co (LSSC). Public reporting, cited by OpenAI, has identified both as alleged task schemes. OpenAI stated it has not independently verified the nature of these entities, but found that messages supposedly coming from them were generated by Chinese-speaking ChatGPT users, likely operating from Cambodia.

The claimed industries spanned stock trading, scooter rentals, and social media engagement sales. OpenAI flagged one telltale absurdity: the network offered to pay more than $5 for a single TikTok like. The company's manual review of online marketplaces found that some sellers of social media likes charge less than $10 for 1,000 likes — meaning the promised rate was roughly 500 times the market price.

Ping, zing, sting

By combining off-platform indicators with internal observations, OpenAI identified a recurring three-stage workflow, which it labeled the ping, the zing, and the sting.

The ping (cold contact). The network generated content for cold outreach, offering unusually high wages for minimal work or high returns on stock-market investments — including high pay for liking social media posts.

The zing (generate enthusiasm). The network translated conversations, likely between operators and their "employees." These exchanges covered task logistics but were frequently interspersed with motivational messages about earnings and potential bonuses.

The sting (extract money). The network generated or sent content pressuring the "employee" or "investor" to contribute money to unlock larger rewards. The pressure took several forms: an initial "deposit," cryptocurrency purchases, and "handling fees."

One cold-call message generated with ChatGPT and distributed by the network was sent to eight phone numbers simultaneously, none of which were in the sender's contact list. One recipient left the group immediately. A Telegram message captured in the report instructed a potential victim to purchase £20 worth of cryptocurrency and transfer it to an unnamed merchant.

Public reporting cited by OpenAI suggests some of the companies operated by charging new recruits substantial joining fees, then using a portion of those funds to pay existing "employees" just enough to keep them engaged — a structure characteristic of task scams as described by the FTC.

Uncertain but real impact

OpenAI cautioned that quantifying the network's true reach is difficult given its limited visibility. But the evidence of real victims exists. Off-platform reports and conversations in which "employees" demanded refunds indicate that at least some individuals paid these alleged employers. OpenAI also observed genuine users defending the companies on social media, suggesting a degree of real-world engagement beyond the scammers' own sock puppets.

The episode closes a loop that platform-security teams increasingly recognize: AI-generated fraud at scale creates AI-detectable patterns at scale. The same centralized operation that let the network coordinate across six languages and multiple platforms also left a consistent footprint across ChatGPT conversations — one that OpenAI could trace, translate, and shut down.