Cambodia scam network: observed model support, unknown AI-attributable loss

OpenAI, “Disrupting a Cambodia-based criminal scam operation,” 31 July 2026. A tip from WhatsApp led investigators to a network of model accounts they assessed as probably operating near Poipet. The accounts drafted/translated conversations for WhatsApp and Telegram targets, developed fictitious profiles and document images and also managed internal staff communications. Operators requested deposits, activation payments and proof-of-payment screenshots. The report also observes messages consistent with coercive recruitment and possible trafficking, while explicitly not establishing the circumstances of individuals.

This connects ordinary service access to a real actor and external targeting workflow, stronger than a lab-crafted phishing email. Yet OpenAI saw one provider’s account interactions and scammers’ own descriptions: it could not verify their claims of victim losses, or measure the number of delivered contacts, conversion, AI’s incremental efficacy, money retained after interdiction, or displacement to other providers after account bans. Its defensive action was account removal and sharing indicators. Combined with IC3’s bank-freeze evidence, this is a plausible scam-scaling mechanism, not a measured aggregate AI-attributable fraud trend. Research brief.