OpenAI Says Russia’s Dark Clark Is Its First Category 5 Influence Op, With a Fake Latin America Think Tank and Nearly 100 Iranian Bylines

OpenAI said on October 8, 2026 that it banned clusters of ChatGPT accounts originating in Russia and Iran after they used the company’s models to run sophisticated “false front” influence operations. The Russia-origin campaign, nicknamed Dark Clark, is the first Category 5 operation OpenAI has disrupted since it began publishing these reports. The Iran-origin campaign, Bogus Bylines, reached Category 4 by planting long-form articles under fake Western journalist personas.

According to OpenAI’s own disruption report, both campaigns combined AI with older tradecraft: fake entities, VPN access, pitch emails, and in the Russian case, unwitting staff on the ground. The company rated impact on the IO Breakout Scale from 1 (lowest) to 6 (highest). Category 5 means OpenAI saw evidence of public comment by politicians in multiple countries, not just social media noise.

That claim is unusually strong for OpenAI’s safety series. Secondary coverage from CyberScoop, NPR, The Hill, CNA, and Bloomberg Law tracks the same figures. OpenAI still did not name a Russian or Iranian intelligence service as the operator.

What OpenAI actually banned

OpenAI says it has exposed 30 covert influence operations over about two and a half years. Most score low. Campaigns that land in real outlets tend to score higher than pure social spam.

Dark Clark and Bogus Bylines fit that pattern. OpenAI compared them to older fronts such as the fake journalist “Alice Donovan” (2016-17) and Meta’s 2020 “PeaceData” site that recruited unwitting writers. Both faded after disclosure. OpenAI’s goal is the same: make continuing the work harder.

Dark Clark: a fake think tank with real employees

OpenAI banned ChatGPT accounts that originated in Russia. Most operators prompted in Russian. One prompted in Spanish but still appeared to be in Russia. Because OpenAI blocks model access from Russia, they used VPNs.

The nickname comes from “Mia Clark,” a fabricated persona OpenAI says the operators used to run social activity and talk to staff for a self-described research platform called the Social Research Center (SRC). On its website, SRC focused on the Indian diaspora in Latin America and India-region relations. Internal reports referenced hiring, firing, pay scales, and research plans, which OpenAI reads as control of the entity, not a loose partnership.

OpenAI says the available evidence indicates Latin American employees did not know they were working for a Russian group. It identified well over 60 articles on the SRC site, most of them original. That complexity, the company wrote, stands out as the most complex front-identity attempt it has disrupted in this reporting period.

The operators’ ChatGPT workload was mostly status reports to an unknown superior. Those memos described three workstreams: damaging Ukraine’s reputation in Latin America and undermining recruitment for Ukraine’s armed forces; interfering in domestic politics, especially in Argentina and Bolivia; and managing the SRC.

Which Dark Clark claims OpenAI could corroborate

OpenAI is careful here, and readers should be too. The operators’ own reports tried to take credit for events they may not have caused. Some claimed fakes never showed up in open sources. Still, several matched public fact checks, media stories, or official denials.

• Peru, May 2026: Fake education emails pushed Ukraine-themed school events referencing Stepan Bandera. OpenAI says matching stories appeared in Peruvian and Polish press, with comment from at least one Polish MEP.
• Ecuador, June 2026: Fake emails allegedly pushed school ceremonies tied to President Daniel Noboa and Erik Prince. OpenAI cites matching coverage and an education minister rebuttal.
• Ecuador, March 2026: Fake audio of Ukraine’s consul and a fake recruitment video. Ecuadorian fact checkers described a matching campaign in early April.
• Bolivia, late May 2026: Fake audio warning that La Paz water would be shut off. OpenAI points to a debunk and an official EPSAS denial.

The operators also asked about a Russia-linked network open-source researchers call Politology or La Compania far more than other Russia-origin groups. OpenAI does not equate Dark Clark with that network, but notes some 2024 Argentina claims in the memos match public attributions to it.

Bogus Bylines: seven personas, nearly 100 articles

The Iran-origin cluster prompted in Persian and produced Persian and English content, again over VPNs. OpenAI nicknamed it Bogus Bylines for seven fake bylines used to pitch outlets: Ervin B. Hoskins, Noah Lamington, Sophia Gonzalez, Michael Harrison, Ericka Feusier, Jenny Williams, and Alice Johnson.

Personas were framed as Western journalists, often American. For Michael Harrison, operators asked ChatGPT to write the biography. OpenAI notes Meta disrupted an Iranian IO featuring a Michael Harrison persona in March 2026, and transparency tools placed the Hoskins accounts in Iran before Meta banned that Instagram account in August 2026.

OpenAI identified almost 100 articles under those bylines across roughly a dozen small to medium outlets on international affairs and the Middle East, from July 2025 to October 2026. Volume rose after the US-Iran conflict began. Typical prompts tuned drafts to an outlet’s rules, then drafted the pitch email.

One outlet that carried the work had almost 2 million Facebook followers as of August 2026, OpenAI said. Social comment batches were weaker: low engagement, many likes from linked accounts, and a self-serving metric that counted views on posts they replied to. OpenAI rates commenting at Category 2 and articles at Category 4.

Overall activity, OpenAI wrote, “appeared consistent with a commercial actor running a for-hire influence campaign.” It could not identify the actor. Information was shared with authorities and industry partners.

Confirmed vs unconfirmed

What it means for Indian developers

SRC’s public brand focused on the Indian diaspora in Latin America. That does not make Indian readers the main target, but it shows how diaspora-facing research brands can supply local cover.

For developers and product teams in India, the US, Canada, and Australia:

• False fronts beat fake follower farms. OpenAI’s matrix of 30 IOs says media placement still beats pure social spam for breakout. If you run a publication, newsletter, or community site, treat unsolicited long-form pitches as an attack surface.
• AI makes the paperwork cheaper, not the whole op. Dark Clark mostly used ChatGPT for reporting up the chain. Bogus Bylines used it for polish and pitches. The hard parts were still humans, emails, and local cutouts.
• Transparency metadata matters. OpenAI leaned on app-store and admin-location clues on X and Instagram. Builders shipping social products should assume investigators will read those fields.

The report also sits beside a wider OpenAI safety debate we have covered, including Sam Altman’s Politico comments on accepting some AI harms, David Robinson’s exit essay, research on AI coding agents that erase their own logs, and SoftBank’s larger OpenAI stake.

The skeptical read

OpenAI is both the platform and the narrator. Category ratings are its own framework. Several of the loudest Dark Clark impacts rest on matching open-source stories to operator memos, not on a public forensic dump of every planted file. Operators also tried to claim credit for Argentina and Brazil narratives that already existed without them.

Even so, the Category 5 label is new for this series, the SRC cutout looks more durable than plagiarized fronts OpenAI has flagged before, and nearly 100 placed articles under fake bylines is a concrete media-integrity problem. Once the fronts are named, editors and platforms have something to search for.

FAQ

What is Category 5 on OpenAI’s scale?

OpenAI’s IO Breakout Scale runs from 1 to 6. Dark Clark is Category 5 because evidence suggested public comment by politicians in multiple countries. OpenAI says it is the first Category 5 op in this series.

Did OpenAI say the Russian or Iranian governments ran these campaigns?

No. It attributed the banned ChatGPT accounts to Russia-origin and Iran-origin operators and said it shared information with authorities. It did not name a specific government agency as the sponsor.

What was the Social Research Center?

A self-styled research platform OpenAI says Russian operators controlled through the fake persona Mia Clark. OpenAI says Latin American staff appear to have produced original research without knowing who sat behind the front.

How many fake journalists did the Iranian campaign use?

Seven named personas. OpenAI says it found almost 100 articles under those bylines across roughly a dozen outlets between July 2025 and October 2026.

Was ChatGPT mainly writing the propaganda?

Not for Dark Clark. Those operators mostly used ChatGPT for internal reports, with occasional scripts and letters. Bogus Bylines used it more directly for article polish, pitches, and comment batches.

Featured image: “OpenAI logo with magnifying glass” by Jernej Furman, via Wikimedia Commons, licensed CC BY 2.0. Cropped and resized.