OpenAI Cuts Off Cambodia-Based Scam Ring Running Frauds Through ChatGPT

Key points

• OpenAI banned a coordinated network of ChatGPT accounts it links to a scam operation based in Poipet, Cambodia.
• The accounts helped run investment fraud, romance scams, illegal gambling promotions, and impersonation of law enforcement.
• OpenAI assesses with medium confidence that the operators sit inside Southeast Asia's well-documented scam compound ecosystem.
• The chatbot was used to translate messages, write persuasive scripts, and manage victim conversations at scale.
• The takedown follows a pattern of OpenAI publicly disrupting state-linked and criminal misuse of its models roughly every quarter.

OpenAI has shut down a cluster of ChatGPT accounts it says were being used to run a sprawling scam operation out of Poipet, a Cambodian border town long associated with industrial-scale online fraud.

The company disclosed the takedown in its latest threat report, first surfaced by The Hacker News. It says the accounts helped criminals write and translate messages for investment scams, romance scams, gambling promotions, and schemes where fraudsters posed as police officers to frighten victims into paying up.

Think of it as a call centre for fraud, with the chatbot acting as a very fast copywriter and translator sitting behind every conversation.

Who was behind the accounts?

OpenAI attributes the activity, with what reads as medium confidence, to a criminal network operating from Poipet in Southeast Asia. The company stops short of naming a specific group, and no vendor tracks this cluster under a formal name the way CrowdStrike or Mandiant would label a nation-state crew.

That caution matters. Poipet and the wider Cambodia-Myanmar-Laos border zone host dozens of scam compounds, many staffed by trafficked workers. Multiple compounds can share infrastructure, tooling, and playbooks, which makes clean attribution to a single operator hard.

What were the scams?

Four overlapping fraud types, according to OpenAI. The chatbot was used to make each one faster and more convincing across languages.

These are the same social-engineering playbooks Interpol and regional police have flagged for years. What is new is the automation. A single operator can now hold fluent conversations in half a dozen languages at once, which used to require a room full of people.

How did OpenAI find them?

Behavioural signals, mostly. The company looks for coordinated patterns across accounts: shared prompts, shared writing tasks, shared working hours, and infrastructure overlaps with previously banned accounts. That is the same tradecraft used to hunt state-linked influence operations, applied here to organised crime.

OpenAI has been publishing these disruption reports roughly every three months. Earlier editions covered accounts linked to North Korea's Kimsuky cluster (as Mandiant tracks it), Chinese groups overlapping with what Proofpoint calls TA415, and Russian information operations. This is the first time a Cambodian scam-compound operation has featured prominently.

Should ordinary people worry?

Yes, but the defence has not changed. The scams themselves are old. Fake investment platforms, sudden online romances that pivot to money, calls from someone claiming to be a police officer demanding a transfer: all of these predate AI.

What AI changes is polish. The broken English that used to give a scam away is gone. Messages now read fluently in Mandarin, Khmer, Vietnamese, Thai, and English, sometimes within the same conversation.

A few practical habits still work. Verify anyone asking for money through a channel you chose, not one they gave you. Assume any unsolicited investment opportunity is fake until proven otherwise. Real police forces do not demand payment over the phone.

Capability is rising. Intent has not changed. The people running these compounds want your money, and now they have better tools to ask for it.