“Dark Clark”: OpenAI says a Russian network used ChatGPT to run a fake think tank and fakes across Latin America

OpenAI says it has disrupted a large covert Russian influence network that it codenamed “Dark Clark”, Radio Svoboda (RFE/RL’s Russian service) reported on 19 October 2026. According to the company, the operation targeted Latin America; OpenAI said it also blocked a separate Iranian operation. The claims below are OpenAI’s, as summarised by Radio Svoboda.

Who was behind it

According to OpenAI, the network was run from Russia through commercial VPNs, and its operators reportedly included suspected officers of Russian military intelligence (GRU). OpenAI linked it to a Russian organisation called “Politologiya”, also known as “the Company” (La Compañía), which Radio Svoboda describes as one of the successors to the late Yevgeny Prigozhin’s structures, including Wagner and the Internet Research Agency. Most operators wrote to ChatGPT in Russian; one wrote in Spanish but was also located in Russia, according to the report.

How AI was used

OpenAI says the operators did not use ChatGPT mainly to mass-produce posts. Instead, the AI served as a secretary, analyst and translator. It was used for three main tasks:

• writing and updating internal reports for an unnamed supervisor;
• polishing propaganda so it sounded locally authentic, including translation into regional dialects of Spanish and Portuguese;
• running a fake “Social Research Center” (CSI).

The centre was fronted by an invented director, “Mia Clark”, after whom the operation is named. According to OpenAI, it hired real Latin American researchers, sociologists and journalists who did not know who was behind it, and published more than 60 original articles on neutral topics such as BRICS expansion. OpenAI called it the most elaborate false-identity front disrupted by major tech companies in the past two and a half years. Radio Svoboda compares it to the Prigozhin-linked “PeaceData” operation that Meta and others took down in 2020.

OpenAI rated the campaign at level 5 of 6 on the “Breakout Scale”, a tool designed in 2020 by investigator Ben Nimmo. OpenAI said it was the first AI-enabled influence operation documented in detail to reach that level, where content is amplified by prominent public figures.

Examples cited

• Peru (May 2026): a spoofed email that looked like it came from Lima’s regional education authority told schools to hold lessons on solidarity with Ukraine and to repeatedly mention Stepan Bandera. Some schools sent back reports with photos. Articles then appeared in Peruvian and Polish media, and in an English-language outlet in Hungary, claiming Ukraine was “exporting ultranationalist ideologies.” A Polish MEP reacted publicly, according to OpenAI.
• Ecuador (June): a fake Education Ministry order told schools to have pupils swear an oath to President Daniel Noboa and to Blackwater founder Erik Prince. The education minister issued a public denial. Other fakes cited include an audio clip attributed to Ukraine’s honorary consul and an AI-made video claiming the government was recruiting young men to fight in Ukraine. Fact-checkers debunked the video in early April.
• Bolivia: fabricated audio of a supposed central bank official warning of cash-withdrawal limits, a forged fuel-restriction letter, and fake audio about water cut-offs in La Paz. OpenAI said these posts were removed or gained little traction.
• Argentina (2024): internal reports describe campaigns against President Javier Milei, including a false story about luxury Cartier collars for his dogs and paid anti-Milei graffiti in Buenos Aires.
• Brazil: an AI-made story about a fictional Brazilian model supposedly trafficked to Kyiv by Ukrainian embassy staff and killed.

OpenAI argues that local hiring, credible front organisations and dialect-accurate AI translation make such campaigns harder to spot than traditional troll farms. Russia’s response to the allegations is not mentioned in Radio Svoboda’s article.