OpenAI has spent the past two and a half years reporting on how threat actors try to misuse its models for cyber attacks, scams, and covert influence operations. In its latest disclosure, it says it banned two recently detected influence operations: a Russia-origin operation it calls “Dark Clark” and an Iran-origin operation it calls “Bogus Bylines.” Both combined AI with traditional techniques to build sophisticated “false front” entities used to launder geopolitical and conflict-related messaging into target audiences and, in some cases, legitimate media outlets.

The Russia-origin operation targeted countries across Latin America, primarily to undermine Ukraine’s reputation and to influence local political outcomes, especially in Argentina and Bolivia. Operators accessed ChatGPT through VPNs and used it mainly to draft internal reports to an unknown superior, create operation content, and write progress reports on a self-described “research platform” called the Social Research Center (SRC). OpenAI says the operators appear to have controlled the SRC through a fake persona named “Mia Clark,” while co-opting employees in Latin America who did not know they were working for a Russian operation. The SRC produced mostly original content, which OpenAI describes as the most complex front-identity attempt it has disrupted in the past two and a half years.

Dark Clark’s internal reporting describes fake stories, fake leaked documents, and fake audio scripts spread across the region. These included a false report that Argentina’s president bought Cartier collars for his dogs, anti-Milei graffiti paid for by local actors, a fake email to schools in Peru instructing them to hold Ukraine-themed events referencing Stepan Bandera, and a fake email tricking schools in Ecuador into holding a pledge ceremony to the president and Erik Prince. OpenAI says open-source research corroborated several of these claims through fact checks, news coverage, and official denials, while other claimed fakes could not be verified. The operators also tried to take credit for real events they had not caused, which OpenAI interprets as an attempt to inflate their value to their own employers.

The Iran-origin operation used seven fake journalist bylines to pitch long-form English-language articles about the US-Iran conflict to small and medium online outlets. OpenAI identified almost 100 published or syndicated articles across more than a dozen outlets. The operators also generated batches of social media comments in English and Persian, but these generally received little engagement and made up only a minority of comments on the target posts. Their internal reports used a deceptive metric based on views of the posts they replied to rather than views of their actual replies, greatly exaggerating their impact. OpenAI assesses the article workstream as Category 4 on the IO Breakout Scale and the social media comment workstream as Category 2.

Using the same scale, OpenAI assesses the Russia-origin operation as Category 5, making it the first Category 5 influence operation OpenAI says it has disrupted since it began this reporting. Dark Clark also resembles earlier pre-AI false-front operations such as the “Alice Donovan” journalist persona and the “PeaceData” fake news outlet, which stopped after being exposed. OpenAI notes that AI can give false-front operations greater scale, efficiency, linguistic fluency, and editorial ability, and that responsible disclosure is a key vulnerability because exposed front entities are harder to keep operating. The company says it has shared information on both cases with relevant authorities.