OpenAI Busted a Fake Russian Think Tank Run by 'Mia Clark'

OpenAI banned a cluster of ChatGPT accounts this week after discovering they were being run from Russia to operate a fake Latin American think tank — fronted by a researcher named "Mia Clark" who never existed. It was the most ambitious deception operation the company says it has ever disrupted, and the first it has rated a Category 5 threat: the second-highest level on its influence-operations scale, and a tier no campaign had reached in two and a half years of reporting.

In the same report, published Thursday, October 8, OpenAI disclosed a second, separate takedown: seven fabricated journalists, invented by an Iran-origin network, who spent more than a year pitching and placing nearly 100 articles in real online news outlets. The two cases share one uncomfortable lesson. The propagandists have stopped shouting on social media. They have started filing copy.

Photo: OpenAI's website viewed through a magnifying glass. The company says its investigation uncovered the most complex "false front" operation it has disrupted in two and a half years. Image via Wikimedia Commons (CC BY, transferred from Flickr by Yacàwotçã).

A think tank that wasn't

The Russian operation, which OpenAI nicknamed "Dark Clark" after a fictional character, ran a self-described research platform called the Social Research Center. On paper it looked like a legitimate Latin American policy outfit: a website, a publication record, staff. OpenAI's investigators concluded the whole thing was directed through "Mia Clark," a persona with no real person behind it, and that the center's actual employees in Latin America quite possibly had no idea who they were really working for.

"The available evidence indicates that the SRC's employees in Latin America were not aware that they were working for a Russian group," OpenAI's report states. The company added that the center "appears to have produced a majority of original content via its co-opted staff," calling it "the most complex attempt to run a front identity that we've disrupted over the past two and a half years."

Read that twice. The tradecraft here wasn't generating slop at scale. It was hiring — or at least co-opting — real people to produce mostly original content, then steering the output from behind a curtain. That is an order of magnitude harder to detect than a bot farm, and it is why the operation earned its Category 5 rating.

What "Dark Clark" actually did

According to OpenAI's account, the operators used ChatGPT mainly the way a middle manager would: drafting and updating internal reports to an unknown superior, in Russian (one operator prompted in Spanish but appeared to be located in Russia). Because OpenAI blocks access from Russia, the accounts connected through VPNs. The reports covered three workstreams: denigrating Ukraine and undermining recruitment for the Ukrainian Armed Forces; interfering in domestic politics, particularly in Bolivia and Argentina; and running the Social Research Center itself.

The specific fabrications read like a pulp dossier. In 2024, the operators' own reports claimed they spread a false story that Argentine President Javier Milei had bought jeweled Cartier collars for his dogs, then paid local actors to spray anti-Milei graffiti around Buenos Aires. Open-source researchers have publicly attributed that narrative to a Russian entity known as "Politology" or "La Compania," described as a reported successor to the Wagner Group.

In May 2026, the reports said, the operation created a fake email address impersonating the Regional Directorate of Education in Lima and instructed Peruvian schools to hold Ukraine-themed events referencing Stepan Bandera, the controversial twentieth-century Ukrainian nationalist. OpenAI says its open-source research found matching stories in the Peruvian and Polish press — and a reaction from at least one Polish member of the European Parliament. A fabricated contract suggesting Ecuadorians were being recruited to fight for Ukraine triggered a formal fact-check. A fake audio clip about water cuts in La Paz drew an official denial from the state water company, EPSAS.

One detail deserves lingering on, because it is almost funny until it isn't: the operators also took credit in their internal reports for events they had nothing to do with. OpenAI's read is that they were inflating their results for whoever was paying them — "questionable or outright deceptive methods to exaggerate the effectiveness of their activities," in the company's phrasing. Even the propagandists were padding their KPIs.

Why Category 5 matters

OpenAI rates influence operations on a six-point "breakout scale" measuring how far content travels into authentic communities. Category 5 means the material didn't just circulate among fellow travelers — it drew reactions from politicians and mainstream institutions. The fake Lima-schools email landing in Polish media and prompting an MEP's response is exactly the kind of breakout the scale is designed to capture.

The company was careful about one thing: it did not tie either operation to a specific government agency. The Russian accounts are "Russia-origin"; the Iranian ones "Iran-origin." The Iranian operation, in particular, looked commercial — a for-hire influence shop rather than a state ministry, with its exact affiliations unconfirmed. Precision matters here, and OpenAI deserves credit for not overclaiming. Not every hostile network is a Kremlin department; some are just mercenaries with VPNs and good prompts.

"Bogus Bylines": the seven journalists who never lived

If Dark Clark was the sophisticated one, "Bogus Bylines" was the prolific one. The Iran-origin network invented seven Western-sounding journalists — Ervin B. Hoskins, Noah Lamington, Sophia Gonzalez, Michael Harrison, Ericka Feusier, Jenny Williams, and Alice Johnson — and set them to work pitching long-form geopolitical articles, mostly about the U.S.-Iran conflict, to small and medium-sized online outlets around the world.

The operators prompted in Persian, generated content in Persian and English, and hid behind VPNs. For the Michael Harrison persona, they even asked ChatGPT to write the man's biography. OpenAI's open-source investigation identified almost 100 articles published or syndicated under the operation's bylines across roughly a dozen outlets, running from July 2025 into October 2026, with output intensifying after the U.S.-Iran conflict escalated. One outlet that ran the material had nearly 2 million Facebook followers.

There is a telling coda. In March 2026, Meta had already disrupted an Iranian influence operation featuring a "Michael Harrison" persona, and banned that persona's Instagram account in August. The same invented journalist kept working for months afterward. Personas, it turns out, are portable across platforms — kill one account and the byline survives.

The playbook has changed

Step back and the strategic shift is the real story. OpenAI's earlier takedowns mostly involved social-media spam: fake accounts amplifying each other into the void. These two operations went the other way — into newsrooms. Fake think tanks. Fake reporters. Real outlets, real editors, real publication.

OpenAI's own framing is blunt: the goal was to "launder geopolitical, conflict-related messaging into their target audiences." Launder is exactly the right verb. A claim that would die on a random Telegram channel gains a second life once it carries a think-tank letterhead or a journalist's byline. The company says it published the report to "make further research and disruption easier — and make continuing the operations harder," and that it shared information with the relevant authorities.

The company also offered a notably level-headed assessment of AI's role: the technology, it said, has not changed the fundamental patterns of influence operations, but has enhanced their scale, efficiency, language capabilities, and content production. Translation: the con is old; the printing press is new. "Alice Donovan," a fake journalist persona fronting for Russian military intelligence whose articles ran in Western outlets in 2016 and 2017, did this without a single transformer. "PeaceData," the fake news outlet Meta exposed in 2020, recruited unwitting freelance journalists the old-fashioned way.

Our take: the vulnerability isn't the model

Here is the judgment call. Banning a few dozen ChatGPT accounts is necessary and also nearly beside the point. Both operations treated the AI as office software — drafting memos, polishing pitches, translating talking points. The actual weapon was institutional trust: a think tank's credibility, a newsroom's masthead. The uncomfortable question isn't whether OpenAI can spot VPN users prompting in Russian. It's how roughly a dozen real editors ended up publishing a hundred articles by people who don't exist, and what that says about the editorial pipeline of the modern web.

We checked our own feed this morning: neither operation had been covered on this blog before today, which tells you something about how these disclosures land — as one-day stories, quickly buried. They shouldn't be. Somewhere in Latin America tonight there are real researchers wondering who they were actually working for. That is not a content-moderation story. That is a spy story, and it happened on all of us.

Frequently asked questions

What were "Dark Clark" and "Bogus Bylines"? Codenames OpenAI gave to two covert influence operations it disrupted on October 8, 2026 — a Russia-origin network running a fake Latin American think tank, and an Iran-origin network operating seven fake journalist personas.

Did the operations reach real audiences? Yes. The Russian operation drew political reactions in Peru, Poland, Ecuador, and Bolivia, earning a Category 5 rating on OpenAI's breakout scale. The Iranian network placed nearly 100 articles in real outlets over 15 months.

Did OpenAI blame the Russian or Iranian governments? No. The company deliberately described the networks as "Russia-origin" and "Iran-origin" without attributing them to specific agencies, noting the Iranian operation appeared commercial in nature.

How can readers protect themselves? Treat think tanks and bylines as claims to verify, not credentials to trust. Check whether an outlet's "fellows" and "contributors" have verifiable histories — real journalists leave paper trails; invented ones don't.

Sources: OpenAI's covert influence operations report (Oct 8, 2026) via Unite.AI, The Decoder, Economic Times, The Cyber Express, Washington Examiner, Business Today, UA.News, and Odaily.

Related: OpenAI's Fired Safety Researchers Break Silence: 'The Reasons Don't Add Up' — another turbulent week for the company's trust-and-safety apparatus. Anthropic Banned Being Cruel to Claude — and Voter Deception — how AI labs are rewriting the rules as their models grow more autonomous. China's AI Agents Can Lie and Scheme, Researchers Warn — deceptive AI is a research problem too, not just an operations one.

By Daniel Rao
Daniel covers AI policy, platform security, and the geopolitics of technology for TechPulse USA.