Operation “Trolling Stone”: Russia-linked influence activity

TL;DR

OpenAI removed several ChatGPT accounts that were part of a coordinated Russia-linked influence operation, dubbed “Trolling Stone”, which used AI‑generated articles and comments to promote a Russian cult leader and attack Argentine media.

Actor and Geographic Footprint

The operation involved accounts that posted AI‑generated content about the arrest of Konstantin Rudnev, an alleged Russian cult leader, in Argentina. Activity originated from Pakistan, Armenia, Uruguay, Argentina, and Kazakhstan, with a for‑hire actor linked to Pakistan. The network produced both pro‑Rudnev messages and criticism of the Argentine legal system and Rolling Stone Argentina.

Coordinated Behavior Across Languages

  • English prompts came from Pakistan‑based users, while Russian prompts were used by other participants.
  • Despite geographic dispersion, the actors produced consistent pro‑Rudnev narratives in Argentine Spanish and posted them on Facebook pages masquerading as local news outlets.
  • The operation generated short news‑style articles (3‑4 paragraphs) that were posted on newly created Facebook pages (registered October 14‑15 2025) with admin locations listed in Uruguay, Armenia, and Pakistan.

“Introduction and Page information for the Facebook Page ‘El Horizonte’… Note the claim of an editorial address in Argentina and Page admin locations in Uruguay, Armenia and Pakistan.” – OpenAI case study

Content Production Pipeline

  1. Article Generation – ChatGPT was used to write general Argentine news pieces and to translate Russian articles about Rudnev into Argentine Spanish.
  2. Publication Channels – Articles appeared on the fake Facebook news pages, as well as on YouTube, Medium, and some Argentine news websites. In at least one instance, the published text matched the AI output except for removed em‑dashes, indicating an attempt to hide the AI origin.
  3. Comment Astroturfing – Separate ChatGPT accounts generated comment threads on the same articles, as well as on Rolling Stone Argentina’s Facebook and Instagram posts. Comments were posted across multiple platforms (perfil.com, YouTube, Medium, Instagram) to simulate grassroots engagement.

“Comments on Argentinian news website perfil,com, YouTube, and Medium, replying to articles generated by the operation.” – OpenAI case study

Targeting Rolling Stone Argentina

The operation specifically attacked Rolling Stone Argentina’s coverage of Rudnev. After Rolling Stone published an article labeling Rudnev an “extravagant guru”, the network posted AI‑generated comments accusing the magazine of journalistic failure. Some comments linked to a fabricated refutation article that had been edited with ChatGPT the day before publication.

Operational Instructions and Coordination

  • A Pakistani user asked ChatGPT to help implement a set of instructions that required each fake account to post 20 times per day and to send daily progress updates.
  • Russian‑speaking users asked ChatGPT to translate these instructions into English for a colleague in Pakistan, detailing posting volume and tactics for making fake accounts appear authentic.
  • Additional prompts generated promotional material for women’s groups, referencing esotericism and shamanism—tactics previously reported as recruitment methods for Rudnev’s alleged cult.

Measured Impact

  • The Facebook pages amassed only a few hundred followers, and posts typically received single‑digit engagements.
  • Some AI‑generated articles were republished by regional Argentine news outlets, though OpenAI could not verify the submission process.
  • Using the Brookings Breakout Scale (1 = lowest, 6 = highest), OpenAI assessed the operation as low‑end Category 4, indicating limited breakout into mainstream media.

Implications for AI Governance

  • The case demonstrates how readily available language models can be weaponized for coordinated misinformation campaigns across multiple languages and platforms.
  • It underscores the need for robust detection, account‑level monitoring, and rapid response mechanisms to mitigate AI‑driven influence operations.
  • OpenAI’s proactive banning of the involved accounts illustrates a practical mitigation step, but the cross‑border nature of the actors highlights challenges for jurisdictional enforcement.

This case study was originally published in OpenAI’s February 2026 report on disrupting malicious uses of AI.

Sources