For a year and a half, a fake Latin American think tank sent letters to education officials in Peru and Ecuador about school events supporting Ukraine that never happened — and almost nobody caught the forgery. ChatGPT helped write those letters, along with scripts for fabricated audio and fake "leaked" documents.
On October 8, OpenAI described how it uncovered this Russia-linked network and gave it the highest score on the company's own threat scale, the Breakout Scale. The scale runs from 1 to 6, and this is the first time in two and a half years of public reports — covering roughly thirty disrupted operations — that OpenAI has assigned the top category.
The network, which OpenAI internally calls Dark Clark, built a fictitious "Social Research Center" and published more than 60 articles through it. Operators recruited unwitting people in Latin America as front authors, and the campaign had already triggered public comments from politicians in several countries — meaning the propaganda didn't just stay on social media, it reached real politics.
Alongside it, OpenAI found an Iranian network called Bogus Bylines: seven invented "journalists," one of whom ran accounts on X and Instagram under the name Ervin B. Hoskins. Between July 2025 and October 2026, roughly a hundred pieces appeared under these bylines across about a dozen real outlets covering international affairs and the Middle East, usually with an antiwar or left-leaning slant. Here ChatGPT's role was narrower — checking drafts against submission guidelines and drafting pitch emails to editors.
The gap between the two networks shows where AI is actually changing the rules of propaganda: not in the volume of text it generates, but in how quickly a fake persona can sound credible enough to slip past a real newsroom's filters. OpenAI banned the ChatGPT accounts tied to both networks and shared its findings with authorities and industry partners. The company is also candid about the limits of its own visibility: what it caught is not a measure of how many similar networks are still running undetected.
