OpenAI Safety has published details on four influence operations that leveraged AI models to generate content for covert activities. These operations, identified as "STORM-2035" (Iran-origin), "Nine—emdash Line" (PRC-origin), "Zero Zeno" (Israel-linked), and "Trolling Stone" (Russia-linked), involved the use of ChatGPT and other OpenAI models to create articles, social media posts, and comments.
The "STORM-2035" operation, detailed in OpenAI's October 2024 report, focused on generating US and UK election content. The "Nine—emdash Line" operation, described in OpenAI's October 2025 report, produced regional influence content concerning the South China Sea, Hong Kong, and US politics. The "Zero Zeno" operation, outlined in OpenAI's May 2024 report, generated anti-Hamas, anti-Qatar, pro-Israel, anti-BJP, and pro-Histadrut content. The "Trolling Stone" operation, featured in OpenAI's February 2026 report, generated comments about an alleged Russian cult leader's arrest in Argentina.
Key Points
- OpenAI banned accounts linked to Iran-origin "STORM-2035" for generating US and UK election content and publishing it across sites.
- The "STORM-2035" operation used OpenAI models to generate long-form English articles (700-900 words) for five websites and short comments in English and Spanish for X and Instagram.
- OpenAI banned accounts linked to PRC-origin "Nine—emdash Line" for creating regional influence content about the South China Sea, Hong Kong, and US politics.
- The "Nine—emdash Line" operation generated English-language social media posts about Vietnam's environmental impact in the South China Sea and criticism of Philippines President Ferdinand Marcos, as well as Cantonese-language posts about Hong Kong's pro-democracy movement.
- OpenAI banned accounts linked to Israel-origin "Zero Zeno" for generating anti-Hamas, anti-Qatar, pro-Israel, anti-BJP, and pro-Histadrut content for X, Facebook, Instagram, websites, and YouTube.
- The "Zero Zeno" operation used OpenAI models to create fictional personas and bios for social media and conduct research into individuals in Israel.
- OpenAI banned accounts linked to the Russia-linked "Trolling Stone" operation for generating comments about an alleged Russian cult leader's arrest in Argentina.
- The "Trolling Stone" operation, with activity originating from Pakistan, Armenia, Uruguay, Argentina, and Kazakhstan, generated short articles about general news in Argentina and translated articles about the Rudnev case from Russian into Argentinian Spanish.
Context
According to OpenAI Safety, these operations represent instances where AI models were used to generate content for covert influence campaigns. The actors sometimes used OpenAI models for basic research, such as seeking tips on social media engagement or identifying niche blogs and forums. In some cases, the operations attempted to create the impression of audience engagement by having fake accounts comment on their own posts.
Why It Matters
These reports demonstrate how AI models can be misused for generating deceptive content in influence operations. Builders and deployers of AI systems must consider the potential for such misuse and implement safeguards to detect and disrupt these activities, as well as understand the evolving tactics of threat actors.
What To Do
- Note the specific dates of OpenAI's reports (October 2024, October 2025, May 2024, February 2026) to track the timeline of these disruptions.
- Review the types of content generated (long-form articles, short comments, persona creation) to understand the range of AI misuse.
- Observe the identified platforms (X, Instagram, Facebook, YouTube, various websites) where these operations distributed content.
- Consider the low engagement levels reported for these operations, indicating that AI-generated content does not automatically guarantee impact.
Keep Exploring
/atlas/claude-family /atlas/gpt-family /techniques/system-user-separation