OpenAI has published a threat report describing two covert influence operations it banned from ChatGPT, one originating in Russia and one in Iran. Both used the company's models as a back-office tool for "false front" entities: organizations and personas that look independent but quietly push state-aligned messaging into real media.
The Russia-origin network, which OpenAI calls "Dark Clark," ran a self-described research platform in Latin America named the Social Research Center, fronted by a fictitious persona called Mia Clark. According to the report, the operators prompted mostly in Russian, reached ChatGPT through VPNs, and spent most of their usage drafting internal reports for an unknown superior. Their campaigns aimed to damage Ukraine's reputation in the region and to sway domestic politics, especially in Argentina and Bolivia. OpenAI says local staff appear to have been recruited without knowing they worked for a Russian group, and it describes the setup as the most complex front identity it has disrupted in two and a half years.
Some of the material the operators claimed credit for has been tied by outside researchers to "Politology," also called "La Compania," which has been reported as a successor to the Wagner Group network. One example in the report involves a fake email, supposedly from Lima's regional education authority, that urged schools to hold events honoring Stepan Bandera; stories about those events later surfaced in Peruvian and Polish media. Another involves a forged contract used to claim that Ecuador's government helped recruit citizens into the Ukrainian army, a claim fact checkers later rejected. On the IO Breakout Scale, which runs from 1 to 6, OpenAI places this operation in Category 5, the first time it has assigned that rating since it began public reporting.
The Iran-origin network, nicknamed "Bogus Bylines," prompted in Persian and produced text in Persian and English. Its operators used seven invented Western journalist identities to pitch long-form articles on the US-Iran conflict to small and mid-sized outlets. OpenAI says it found almost 100 articles under those bylines across more than a dozen publications, with the earliest from July 2025 and the latest from October 2026. One outlet that ran the pieces had close to 2 million Facebook followers. The same group generated batches of social media replies, but those drew little engagement, so OpenAI rates the article work at Category 4 and the commenting at Category 2. It believes a commercial influence-for-hire actor was behind it but could not identify which one.
A recurring detail is how both groups used AI to overstate their own success. The Iranian team measured impact by counting views on the posts it replied to rather than on its replies, and the Russian operators claimed credit for activity they did not appear to carry out. OpenAI says it has shared information on both cases with relevant authorities and industry partners.
The broader pattern matters for newsrooms and platform defenders. Across the 30 covert operations OpenAI says it has exposed since early 2024, those that placed content in real outlets reached the largest audiences. AI did not invent these tactics, which echo older fronts such as the fake journalist "Alice Donovan," but it made them cheaper to staff and easier to polish.
The findings come from OpenAI's own investigation, and parts of the impact assessment rest on open-source matching that the company acknowledges is incomplete. The attribution to Russia and Iran reflects where the operators were located and how they behaved, not a confirmed link to any named government agency.