In short: OpenAI identified and banned a coordinated network of ChatGPT accounts operated by a group based in Cambodia to run romance scams, fake cryptocurrency investments, gambling fraud, and law enforcement impersonation. The AI was used to create synthetic personas, translate conversations with victims, and generate promotional content at scale. Some of the operators running these schemes were themselves victims of human trafficking and forced labour.
What Happened
On July 31, 2026, OpenAI published a formal threat intelligence report and banned a coordinated cluster of ChatGPT accounts that were very likely operating from Poipet, Banteay Meanchey Province, Cambodia — a region that researchers and human rights organisations have repeatedly linked to online scam compounds and forced-labour operations.
The investigation began with a tip from Meta (owner of WhatsApp), through which the group made initial contact with victims. OpenAI tracked account behaviour, identified the coordinated abuse pattern, and shared intelligence with international law enforcement. All identified accounts have been permanently banned from OpenAI services.
If you receive unexpected outreach on messaging apps — especially from someone who quickly turns the conversation to crypto investments or an online romance — this is the exact pattern described in OpenAI's report. Our security blog covers practical steps for protecting your data during online interactions.
How the Scheme Worked: Ping, Zing, Sting
The group used the classic three-stage fraud sequence known as ping-zing-sting:
- Ping (first contact). Victims were approached on WhatsApp, Telegram, or social media. The opener was typically a "wrong number" text, a message from a "mutual friend" the victim could not recall, or an unsolicited introduction from an apparently successful investor.
- Zing (trust-building). Weeks — sometimes months — of friendly or romantic conversation followed. The victim had no reason for suspicion; the persona was attentive, consistent, and personal. ChatGPT allowed operators to maintain dozens of simultaneous "relationships" across multiple languages without losing the narrative thread.
- Sting (the strike). The scammer introduced a "guaranteed" investment opportunity in cryptocurrency or spot gold, or demanded payment of a supposed fine issued by police or customs for "illegal transfers." Victims transferred funds voluntarily, under the weight of emotional trust built over weeks. Forged screenshots of fake transaction confirmations were provided as false proof.
ChatGPT as an Industrial-Scale Fraud Tool
The central finding from OpenAI's report: AI removed the main operational constraint of romance fraud — scale. Before language models, running convincing simultaneous conversations with hundreds of victims across multiple languages required large teams. ChatGPT eliminated that bottleneck:
- Synthetic identity creation. Accounts generated detailed personas — backstory, career, interests, lifestyle — that held together under extended conversation.
- Message translation and localisation. Cambodia-based operators communicated with victims worldwide in their native languages without requiring human translators.
- Promotional content production. Polished marketing materials for fictitious crypto platforms and gambling sites were generated automatically.
- Internal operations. ChatGPT was used to draft fake job postings, employment contracts, and internal announcements — documents used to recruit new operators into the compound under false pretences.
As early as February 2026, OpenAI disclosed details of "Operation Date Bait" — an earlier iteration of the same group, which used AI to generate images of fake women, logos, and banners for a service called Klub Romantis. OpenAI estimated that operation alone was defrauding hundreds of new victims every month.
The Human Trafficking Layer: Operators Are Also Victims
One of the most troubling dimensions of the investigation concerns the operators themselves. Threat intelligence confirmed that many of the people directly conducting victim conversations were themselves victims of human trafficking.
The recruitment model is consistent across Southeast Asian scam compounds: online job ads promise well-paid tech or marketing work in the region. Applicants travel to Cambodia, where their passports are confiscated, they are moved to guarded facilities, and they are coerced into working as scam contact-centre operators under threat of violence. Debt bondage — charges for accommodation, food, and "training" — prevents escape.
This creates a double chain of victims: those who lose money to the schemes, and those who are forced to run them.
How to Recognise the Scheme and Protect Yourself
Romance fraud and pig-butchering are uniquely dangerous because they attack trust, not technology. Warning signs to watch for:
- An unsolicited first message — "wrong number," "mutual friend you cannot place," or a cold introduction on a professional platform;
- Rapid emotional acceleration: within a week, the conversation feels unusually intimate or the person declares strong feelings;
- The contact never appears on live video, or only on a pre-recorded clip;
- The conversation gravitates toward investment "opportunities" — typically cryptocurrency or precious metals;
- A request to "unlock" already-earned profits by depositing additional funds.
Practical steps that reduce digital risk:
- Verify identities. Reverse image search (Google, Yandex Images), and a live video call are basic tools that a scripted AI persona cannot pass.
- Never send money to someone you have only met online, regardless of how much trust has been built through conversation.
- Protect your data and digital footprint. LiMP VPN encrypts your internet traffic and masks your real IP address, making it harder for bad actors to profile and target you based on your browsing activity. See our features page for how the technology works.
- Report suspected trafficking. If you believe someone is being coerced into running scams, contact the International Justice Mission (ijm.org) or local law enforcement.