OpenAI Bans ChatGPT Accounts Behind "Trolling Stone" Influence Operation

OpenAI has banned a cluster of ChatGPT accounts that ran a coordinated influence operation defending Konstantin Rudnev, an alleged Russian cult leader arrested in Argentina in April 2025. The company nicknamed the network "Trolling Stone" after its most distinctive target: the Argentinian edition of Rolling Stone magazine, which had reported on Rudnev's case.

The case study appeared in OpenAI's February 2026 report on disrupting malicious uses of AI. It describes a geographically dispersed operation whose participants behaved in a coordinated way despite originating, in OpenAI's assessment, from Pakistan, Armenia and Uruguay, with further activity likely from Argentina and Kazakhstan. The Pakistan-origin activity was connected to a for-hire actor.

Rudnev was arrested in April 2025 in southern Argentina along with 14 others, according to Argentinian media reports. El País described him as a Russian cult leader linked to human and drug trafficking. Rolling Stone Argentina's June article called him an "extravagant guru" and the founder of an "exotic and disturbing" cult.

The operation used ChatGPT to generate batches of social media comments about the case. Some supported Rudnev. Others criticized the Argentinian legal system and trolled Rolling Stone Argentina directly.

Fake news outlets and astroturfed engagement

One main workstream produced short articles, typically three or four paragraphs, about general news in Argentina. OpenAI identified these articles on a set of Facebook Pages posing as Argentinian news outlets. All six Pages were registered on October 14-15, 2025. Four of the six displayed admin locations under Facebook's Page transparency settings: three were managed from Uruguay, Armenia and Pakistan, and a fourth from Pakistan and Uruguay.

One such Page, "El Horizonte" ("the horizon"), claimed an editorial address in Argentina while its Page transparency data showed administrators in Uruguay, Armenia and Pakistan.

Alongside the general news output, the operation translated articles about the Rudnev case from Russian into Argentinian Spanish. Some were published on the fake news Pages. Others appeared on Facebook, YouTube and Medium accounts dedicated to the Rudnev case, and a few landed on genuine Argentinian news websites.

On at least one occasion, OpenAI found, the published version was almost identical to the ChatGPT-generated text — except the em-dashes had been removed, suggesting a deliberate attempt to obfuscate the text's AI origin.

Further ChatGPT accounts in the network then generated comments about the articles, and a range of accounts posted those comments on the relevant sites. OpenAI described this as typical "astroturfing" behavior, aimed at manufacturing the impression of genuine grassroots support.

Targeting Rolling Stone

Most of the operation's comments responded to articles the operation itself had generated. The key exception was a set of replies targeting Rolling Stone Argentina. After the magazine's June article on Rudnev, operation accounts generated comments accusing Rolling Stone of failing in its journalistic duties. Those comments were posted under Rolling Stone Argentina's Facebook and Instagram posts about music — including a November 3, 2025 Instagram post featuring cover artist Santiago Motorizado.

One notable detail: some comments criticizing the magazine referenced a web article that apparently refuted its reporting. That "refutation" carried the byline of an apparently Spanish-speaking journalist, but OpenAI determined it had been edited with ChatGPT the day before publication — on October 26, for an article published October 27 — by a Russian-speaking user inside the operation.

Operational instructions routed through ChatGPT

While most of the operation's prompts concerned content generation, a small portion dealt with coordination. On one occasion, the Pakistan-based user asked ChatGPT to help implement a set of operating instructions they had received. The instructions required each fake account to post 20 times per day and demanded daily progress updates. On another occasion, a Russian-speaking user asked ChatGPT to translate a different set of instructions into English, apparently for a colleague of the original user in Pakistan. Those instructions covered expected posting volumes and how to make the fake accounts look convincing.

The Russian-language accounts also asked ChatGPT to generate promotional material for women's groups, including cold-outreach messaging. Some of that material referenced esotericism, shamanism, yoga and meditation — activities that open-source reporting, including from Russian outlet Izvestia and CNN, has described as recruitment tactics into Rudnev's alleged cult. OpenAI said it could not independently confirm the nature of the activities to which the material referred.

Low actual impact, but real placement

As with influence operations OpenAI has documented in earlier threat reports, "Trolling Stone" seemed designed to create the appearance of online engagement rather than actually achieving it. The Facebook Pages typically had a few hundred followers, and typical posts received single-digit engagement.

OpenAI assessed the operation on the Brookings Breakout Scale, which rates influence operations from 1 (lowest) to 6 (highest). The company placed "Trolling Stone" toward the low end of Category 4 — "breakout to mainstream media" — because some of its generated articles were published by regional news outlets in Argentina. OpenAI could not independently confirm how those articles were submitted and accepted.

That placement is the operation's most consequential feature. The engagement metrics were negligible, but the pipeline ran from ChatGPT prompts in Russian and English to translated Spanish-language articles published under fake bylines on Argentinian news sites, then back to AI-generated comments amplifying those same articles. For publishers and platforms, the case shows how little actual audience an operation needs before its content starts crossing into genuine media ecosystems.