OpenAI has reported that over the past two and a half years it observed threat actors attempting to use its models for cyber attacks, covert influence operations (IO), scams, and other policy abuses. Recently the company banned two sophisticated operations—one originating in Russia and the other from Iran—that combined AI with traditional techniques to support false-front entities and place content in mainstream outlets.
Both campaigns used multi-faceted tactics: long-form articles, coordinated social-media comments, fabricated leaked documents and audio, and AI-drafted internal reports. The Iranian operation ran seven fake journalist bylines to pitch and place long articles; the Russian operation appears to have co-opted unwitting local personnel in Latin America to staff a front think tank.
Why these cases matter
- AI made existing influence workflows faster and more linguistically fluent, enabling operators to exploit unsuspecting editors, employees, and audiences.
- Both campaigns succeeded in getting content into actual media outlets rather than relying solely on social media, increasing potential reach and impact.
- OpenAI assessed the Russian-origin operation at Category 5 and the Iran-origin operation at Category 4 on its IO Breakout Scale (1–6).
Russia-origin operation: "Dark Clark" in Latin America
OpenAI banned a cluster of ChatGPT accounts that originated in Russia and accessed services via VPNs. The cluster targeted multiple Latin American countries with campaigns largely aimed at undermining Ukraine’s reputation in the region and at influencing local politics in countries such as Argentina and Bolivia. Most operators prompted in Russian; one used Spanish but also appeared to be located in Russia. OpenAI shared findings with relevant authorities.
Main uses of ChatGPT
- drafting and updating internal reports on campaign activities and claimed successes;
- creating content for operations in some cases; and
- drafting reports on a so-called Social Research Center (SRC), a self-described research platform operating in Latin America.
Key unusual features
- Co-opted local staff: Internal reports indicate the operators managed hiring, pay, and firing decisions at the SRC, suggesting control rather than mere cooperation. Evidence suggests local employees were unaware they were working for a Russian operation.
- Wide penetration: Some fakes tied to the operation provoked fact checks and official denials. Based on open-source corroboration of several claims, OpenAI rated this operation Category 5 — the first Category 5 it has disrupted since it began reporting.
Illustrative tactics and incidents (2024–2026)
- The operators claimed responsibility for two Argentina-targeted campaigns in 2024: a false story that President Javier Milei had purchased Cartier collars for his dogs, and another alleging they paid actors to paint anti-Milei graffiti. Leaked documents have linked similar claims to an entity called “Politology” or “La Compania.”
- In May 2026 they reported creating a fake Regional Directorate of Education email in Lima instructing schools to hold Ukraine-themed events on the Day of Cultural and Linguistic Diversity (May 21). According to the operators, some schools replied and provided photos. Open-source searches found media stories in Peruvian and Polish press and an English-language outlet in Hungary that matched the claim; some articles were later deleted.
- In Ecuador, operators claimed to have spread fake audio attributed to Ukraine’s consul and a forged video alleging government recruitment for Ukraine; fact-checkers and official rebuttals matched this pattern.
- In Bolivia, at the height of protests in late May, the operation spread a false audio claiming the state water company would cut water to La Paz; fact-checkers and official denials addressed these claims.
SRC content and reach
Open-source records showed over 60 articles on the SRC website, most appearing to be original. This distinguishes Dark Clark from other Russia-linked think-tank fronts that largely plagiarized content. The operation also attempted to build a social-media presence but encountered account restrictions and location inconsistencies in platform transparency settings (for example, some SRC accounts showed admin locations or app-store connections tied to Russia or Venezuela).
AI in support roles
The principal role of ChatGPT in this operation was producing polished internal reports that documented covert campaigns. On a smaller scale, operators used the model to translate, adapt, or proofread content (for example, translating a Russian-language letter to Spanish and creating audio scripts posted alongside a matching image and voiceover on TikTok).
Assessment and caveats
OpenAI cautions that operators’ internal claims of impact may be exaggerated; some alleged fakes could not be corroborated in open sources because they were taken down, spread in private channels, or never posted. Nevertheless, several claims matched widely circulated articles and prompted official responses. Based on these confirmations and political reactions in multiple countries, OpenAI placed the Russia-origin operation in Breakout Scale Category 5.
Iran-origin operation: "Bogus Bylines" placing long-form articles
OpenAI also banned a cluster of ChatGPT accounts originating in Iran. Operators prompted in Persian and produced both Persian- and English-language content, obscuring location via VPNs. They used AI to refine long-form articles and editorial pitches, generate batches of social-media comments, and polish internal reports.
Structure and tactics
- The campaign employed seven fake bylines: Ervin B. Hoskins, Noah Lamington, Sophia Gonzalez, Michael Harrison, Ericka Feusier, Jenny Williams, and Alice Johnson. These personas were presented as Western journalists and were sometimes backed by social-media accounts.
- Open-source investigations identified nearly 100 articles published or syndicated under these names across roughly a dozen small-to-medium outlets focused on international affairs and the Middle East. The earliest identified article dates to July 2025; the latest to October 2026. Publication frequency increased after the outbreak of the US–Iran conflict.
- Operators also generated batches of English- and Persian-language replies to social-media posts, often aligned to portray the US as the aggressor and Iran as resilient. Many of these replies showed limited engagement and in several cases were posted by accounts whose transparency settings pointed to Iran.
Impact differences across workstreams
- Article placement: The fake journalist personas achieved measurable success getting long-form articles into multiple outlets, some of which have substantial followings (one outlet had nearly 2 million Facebook followers, ~355,000 on X, and ~544,000 on Instagram as of August 2026). OpenAI rated this aspect Category 4 on the Breakout Scale.
- Mass commenting: The reply batches tended to have low visibility and few interactions; internal reports by operators used a misleading metric (counts of views on the original posts rather than on the replies themselves) that inflated claimed impact. OpenAI rated this activity Category 2.
Operational details
Operators used ChatGPT to: assess drafts against specific outlets’ submission criteria and propose edits; generate pitch emails under the fake bylines; create comment batches in response to screenshots or posts; and format internal statistics into polished presentations. Some of the personas’ social accounts were later suspended.
Final observations
OpenAI emphasizes that while AI can streamline certain workflows used in covert influence campaigns, many tactical elements—fake personas, front organizations, co-opted local participants, and placement in real media—are continuations of pre-AI methods. The company says responsible disclosure and coordinated disruption make continuing such operations harder and aid research into their vulnerabilities. It has shared findings with relevant authorities and industry partners to support further investigation and mitigation.
