AI Safety
OpenAI Bans ChatGPT Accounts Behind Russia, Iran Influence Ops
Oct 9, 2026·7 min read·By Rahish Saifi
On October 8, 2026, OpenAI published “Disrupting AI-enabled false front operations”, detailing two covert influence campaigns it disrupted by banning clusters of ChatGPT accounts from Russia and Iran. Both used classic “false front” tradecraft—fake journalists, a think tank with unwitting staff, forged documents and audio—to launder geopolitical messaging into real audiences. ChatGPT’s role was often backstage: drafting internal reports, polishing pitches, and occasionally helping shape fakes—but the endgame was placement in mainstream and niche media, not viral posts on fake social accounts.
OpenAI says it shared information on both cases with relevant authorities. Operators in both campaigns used VPNs: the Russian cluster because OpenAI does not allow access from Russia, and the Iranian cluster to obscure its location.
Why false fronts still win on reach
OpenAI frames these operations as pre-AI influence playbooks with AI sprinkled in for scale, fluency, and editorial polish. The Iranian campaign’s seven Western “journalist” personas echo historical fronts such as Russia’s Alice Donovan (2016–17) and the Internet Research Agency–linked PeaceData outlet (2020), which co-opted unwitting writers. What stands out in OpenAI’s data on 30 covert IO cases since early 2024 is distribution: campaigns that land stories in external publications tend to score higher on the IO Breakout Scale (Ben Nimmo, Brookings, 2020; 1 lowest–6 highest), which OpenAI uses to rate operations, than those relying on operation-run sites or social promotion.
The October 8 post argues that AI makes those workflows easier—more linguistic bandwidth, faster internal reporting, better pitch emails—while the vulnerability is still human: editors, researchers, and employees who trust a plausible byline or employer brand.
Russia: Operation “Dark Clark” (Category 5)
OpenAI banned a Russia-origin cluster that prompted mainly in Russian (one account in Spanish but apparently still in Russia). Activity targeted Latin America, with goals that included undermining Ukraine’s reputation and influencing domestic politics—especially in Argentina and Bolivia.
The centerpiece was the Social Research Center (SRC), a self-described research platform focused on the Indian diaspora and India–Latin America relations. Internal ChatGPT reports suggest operators controlled the SRC via a fake persona, “Mia Clark”—hence the nickname Dark Clark. OpenAI says on-the-ground SRC employees appear not to have known they worked for a Russian group; the SRC published well over 60 mostly original articles, a more durable front than a recent Russia-linked think tank OpenAI exposed that largely plagiarized content.
ChatGPT’s heaviest use was internal reporting to an unknown superior: tactical updates on denigrating Ukraine, interfering in local politics, and managing the SRC. Operators also claimed credit for fakes that open-source work tied to entities reported as “Politology” or “La Compania”, described in public reporting as a successor network linked to Yevgeniy Prigozhin’s ecosystem.
OpenAI corroborated several high-impact claims. One involved a fake Peruvian education email instructing schools to hold Ukraine-themed events referencing controversial nationalist Stepan Bandera; media coverage followed in Peru, Poland, and elsewhere, and a Polish MEP reacted publicly. Other claimed fakes in Ecuador and Bolivia drew fact checks and official denials. Occasional ChatGPT use supported operational content—for example accent- and institution-aware fakes, translated letters, and audio scripts posted to low-reach social accounts mimicking local news brands.
On the Breakout Scale, OpenAI rates Dark Clark at Category 5—the first Category 5 operation it says it has disrupted since it began publishing IO reports—with evidence of political comment across multiple countries.
Iran: Operation “Bogus Bylines” (articles: Category 4; comments: Category 2)
The Iran-origin cluster prompted in Persian and generated Persian and English content. OpenAI nicknamed it Bogus Bylines after seven fake Western journalist personas used to pitch long-form work:
- Ervin B. Hoskins
- Noah Lamington
- Sophia Gonzalez
- Michael Harrison
- Ericka Feusier
- Jenny Williams
- Alice Johnson
Typical workflow: refine an English draft against a target outlet’s criteria, approve edits, then generate a pitch email. OpenAI identified almost 100 articles published or syndicated under those bylines across roughly a dozen small-to-medium outlets focused on international affairs, geopolitics, and the Middle East—often from antiwar or progressive angles. Coverage concentrated on the US–Iran conflict; articles ran from July 2025 to October 2026, and frequency rose sharply after the US-Iran conflict broke out (dense from March 2026). Some personas were backstopped on social platforms; transparency settings on accounts linked to Ervin B. Hoskins showed Iran despite bios claiming a US freelance identity.
In parallel, operators asked ChatGPT for batches of social comments in English or Persian—often replies to posts about Iran, the US, and Israel, with prompts seeking narratives that framed the US as aggressor and Iran as resilient victim. OpenAI found limited breakout: replies were usually a minority of thread activity, with low engagement; many of the likes came from accounts linked to the operation itself. Internal “impact” slides used a misleading metric (views on the parent post, not the reply), inflating apparent success.
Article placement rates Category 4 on the Breakout Scale; the comment campaign Category 2. Activity looked consistent with a commercial for-hire influence actor, but OpenAI says it could not identify the specific sponsor.
What defenders should take away
Neither campaign was “AI-generated spam on X” as the primary strategy. Both invested in identities and institutions that could survive a cursory Google search, then used ChatGPT to run the back office—status reports, pitch polish, comment farms, and selective content support. OpenAI’s disclosure goal is familiar from earlier exposes: false fronts often shut down after exposure, and naming tactics (VPN access, deceptive metrics, unwitting employees) helps platforms, newsrooms, and researchers harden vetting.
For builders shipping chat products, the case is a reminder that abuse detection must cover long-horizon, low-volume prompting (internal reports) as well as bulk generation—and that the highest-impact misuse may leave the model’s UI entirely once a story is planted with a real editor.