OpenAI disclosed that it recently banned two ChatGPT accounts involved in covert influence operations, one originating from Russia and the other from Iran. Both used models in conjunction with traditional methods to operate fake fronts, injecting geopolitical and conflict information into target audiences.

The Russian operation, named Dark Clark, was rated as a level 5 on the influence scale, which ranges from 1 to 6. This is the first level 5 operation interrupted since OpenAI began disclosures. The account originated from Russia, primarily using Russian prompts, with one using Spanish but still showing a location in Russia. Due to the platform not providing access to Russia, the operators connected via a virtual private network. The company has referred this case to relevant authorities.

This group of accounts targeted multiple countries in Latin America. The content was largely used to damage Ukraine's reputation in the region and also involved local politics, particularly in Argentina and Bolivia. The operators used the model to write internal activity reports, generate content, and draft a performance report for a self-proclaimed research platform called Social Research Center, controlled by a fake identity named Mia Clark. OpenAI stated that the Latin American employees appeared unaware they were working for a Russian group. The company described this as the most complex front identity in the past two and a half years. Public reports indicated that the related content had deceived schools, political figures, and media outlets.

The Iranian operation, named Bogus Bylines, was rated as level 4. The account originated from Iran, using Persian prompts to generate content in both Persian and English, while masking its location with a virtual private network. The operators maintained seven fake journalist identities, submitting long articles to small and medium-sized media outlets globally. The publicly listed names included Ervin B. Hoskins, Noah Lamington, Sophia Gonzalez, Michael Harrison, Ericka Feusier, Jenny Williams, and Alice Johnson.

What is unusual about both cases is that the models were not just writing posts. The Russian side used it to manage ground operations and unaware employees, while the Iranian side used it to maintain a submitable journalist persona. The bans cut off the accounts and generation processes, but already published articles and ground employees are not under the platform's authority.

The buyers are normal users still purchasing model and interface usage, while the sellers are the banned covert operation accounts. The incident was driven by abuse disclosures, not product price reductions. The funding has no publicly known transaction flow, and the cut-off is from restricted areas accessing via virtual private networks. Beneficiaries include platforms receiving leads and alerted small and medium media in Latin America and globally, while those under pressure are operators relying on fake think tanks and fake journalists to manipulate information. Once articles are published, banned accounts cannot retrieve the pages.

Source: Public Information