Skip to content

OpenAI

OpenAI’s first Category 5 op hired unwitting Latin Americans

7 min read

Written by an AI model from the linked sources and checked against them by a second model. How we work

All about: OpenAI models and products →
Image: openai.com

OpenAI has banned the ChatGPT accounts behind two covert influence operations, the company said in an October 8 report. One was Russian and ran a Latin American “think tank” whose local staff apparently didn’t know who employed them. The other was Iranian and placed almost 100 articles under seven fake journalist bylines. The Russian operation, which OpenAI nicknamed “Dark Clark,” is the first it rates Category 5 on the IO Breakout Scale (1 lowest, 6 highest), and the rating rests on traces anyone could find in public, such as fact checks, official denials and a Polish MEP’s reaction; the operators’ own reports count for little because they claimed credit for events they had nothing to do with.

OpenAI says both operations closely resembled complex influence campaigns from before AI and used ChatGPT to make some of the workflows easier. For the Russian team, the main use was drafting and updating internal reports to an unknown superior. In most cases OpenAI saw no campaign content made with its models. Both teams connected through VPNs, and OpenAI has shared information on both cases with the relevant authorities.

What put Dark Clark at Category 5

In May, the operators said, they sent schools in Lima a fake email that appeared to come from the Regional Directorate of Education. It told the schools to hold events dedicated to Ukraine on May 21, Peru’s Day of Cultural and Linguistic Diversity, and to reference Stepan Bandera, a twentieth-century Ukrainian nationalist associated in Russia and Poland with fascism, wartime collaboration and atrocities. According to the operators, some schools wrote back to confirm the events and sent pictures. OpenAI then found matching stories in the Peruvian and Polish press and in an English-language publication in Hungary. Some of those stories quoted a Polish MEP who proposed that people showing “anti-Polishness” be declared persona non grata in Poland.

A second fake email in June led schools in Ecuador, by the operators’ account, to hold a ceremony pledging allegiance to President Daniel Noboa and to Erik Prince, the former head of Blackwater. Coverage in the Ecuadorian press closely resembled that claim, and Ecuador’s Minister for Education issued a detailed rebuttal. In Bolivia in late May, at the height of anti-government protests, the team spread a fake audio clip of a worker at the state water company EPSAS saying the government would cut off water to La Paz. That clip drew a debunk and an official denial from EPSAS.

Two older claims tie the team to a known network. The operators said that in 2024 they spread a false report that Argentina’s President Javier Milei had bought Cartier jeweled collars for his dogs, and paid local actors to put up anti-Milei graffiti in Buenos Aires. Based on leaked documents, both campaigns had already been publicly attributed to “Politology” or “La Compania,” which open-source researchers describe as a successor to the Wagner Group and other entities founded by Yevgeniy Prigozhin. The operators also asked ChatGPT about public reporting on Politology and Wagner far more often than about any other Russian network.

Some of what the reports claimed was simply borrowed. The operators asked the model to find incidents they could take credit for. One was Argentina’s Foreign Minister making two arguments about the Falklands/Malvinas at a UN committee meeting, though both have been part of Argentina’s official position for years. OpenAI says this suggests an attempt to run an influence operation against the operators’ own employers.

ChatGPT played a small part in making the fakes, possibly because the team appears to have included a native Spanish speaker from Latin America. It translated a Russian script into Ecuadorian Spanish and helped draft a letter that implicated Ukraine’s honorary consul to Panama in corruption. In March it also proofread a fake contract between “International Security Solutions FZE” and an individual in Ecuador, down to which words should be in bold. A picture of that contract later circulated as evidence that Noboa’s government was tricking Ecuadorians into joining the Ukrainian army. Fact checkers found that the company does not appear in the registry of Ecuador’s Superintendency of Companies, Securities, and Insurance.

A think tank with real employees

The front was the Social Research Center, a self-described “research platform” that, according to its website, studies the Indian diaspora in Latin America. The operators’ reports covered pay scales, hiring, firing and staffing plans, which suggests they controlled the center and weren’t just working alongside it. They also said they had created a persona, “Mia Clark,” to run its social media and deal with staff on the ground. Those employees interviewed experts across the region and wrote research papers, including one comparing Brazil’s economy under Jair Bolsonaro and Luiz Inácio Lula da Silva. The available evidence suggests they had no idea they were working for a Russian operation.

OpenAI found well over 60 articles on the center’s site, most of them apparently original. A Russia-linked think tank it exposed recently had plagiarized most of its content. OpenAI calls the center the most complex front identity it has disrupted in two and a half years, and compares it to PeaceData, the 2020 fake outlet that recruited unwitting journalists and that Meta exposed. Promotion went less well. The center’s main X account showed a German location but a connection through the Russian Federation App Store. As of August 17, a LinkedIn page with the center’s branding had 961 followers and claimed 200-500 staff, but listed only two employees.

Seven bylines and a pitch email

The Iranian operation, nicknamed “Bogus Bylines,” followed a routine. The operators gave ChatGPT an English-language draft on the US-Iran conflict, asked it to check the draft against a specific outlet’s submission criteria and suggest edits, and then had it write the pitch email to that outlet’s editor. The bios that ran with the articles presented the authors as Western journalists, usually American. Social media transparency settings placed the accounts of one persona, “Ervin B. Hoskins,” in Iran. Meta disrupted an Iranian operation in March that also used a “Michael Harrison” journalist persona, and OpenAI’s operators asked ChatGPT to write that persona’s bio.

The outlets were small to medium sites focused on international affairs, geopolitics and the Middle East, some of them antiwar or progressive. One had almost 2 million followers on Facebook, almost 355,000 on X and over 544,000 on Instagram as of August 2026. OpenAI rates the article work at Category 4 because it consistently broke out into a number of media outlets. Batches of generated social media comments got likes, views and replies only in the single or double digits, and many of the likes came from the operation’s own accounts. That part rates Category 2. OpenAI says the overall activity looks like a commercial, for-hire campaign, though it couldn’t identify the actor.

The team measured its comment campaign with a generous yardstick. It counted views on the posts it replied to, before and at some unspecified time after posting, and ignored views on the replies themselves. That credited the operation with audiences the original posts drew on their own, and OpenAI says the method greatly exaggerated the impact.

The template is older than ChatGPT. Alice Donovan, a fake freelance journalist who fronted for Russian military intelligence, published at least thirty pieces between February 2016 and mid-October 2017, according to a tally by CounterPunch, which ran five of them. Between July 2025 and October 2026, Bogus Bylines’ seven personas produced almost 100.

Why OpenAI publishes

OpenAI has exposed 30 covert influence operations since early 2024. In that set, the ones that get content into real media outlets tend to have the highest potential reach and impact. Both of these operations did that, although not all of the content they placed came from OpenAI’s models. Alice Donovan and PeaceData both stopped after they were exposed, and OpenAI says making these operations harder to continue is the reason it reports them. The latest article OpenAI found under one of the Iranian bylines was published in October 2026.

No comments yet. Be the first.

Join the conversation

Sign in to leave a comment. It appears after moderation.