OpenAI Disrupts Operation STORM-2035 Iranian Influence Activity
OpenAI has disrupted an Iranian-origin influence operation, identified as Storm-2035, by banning ChatGPT accounts used to generate deceptive political content across multiple platforms. This operation highlights the use of AI to scale the production of both long-form articles and short-form social media engagement to simulate diverse political viewpoints.
Identification and Attribution of Storm-2035
OpenAI identified and banned a set of ChatGPT accounts that were generating English and Spanish content for a coordinated influence campaign. The actor's activity was linked across multiple platforms, including more than a dozen fake personas on X, one account on Instagram, and five separate websites.
Attribution was achieved through cross-industry collaboration:
- Microsoft had previously reported some of the websites associated with the actor under the name "Storm-2035."
- Meta confirmed that the identified Instagram account was connected to an Iranian network that Meta had disrupted in December 2022.
OpenAI's visibility into the actor's behavior enabled the first-time connection between the specific social media posting activity and the associated websites.
Operational Workstreams and Tactics
The actor employed two parallel workstreams to disseminate content and build fake personas.
Long-form Article Generation
The actor used OpenAI models to produce English-language articles, typically ranging from 700 to 900 words. These were published on five websites that posed as either progressive or conservative to target different ideological audiences. A primary focus of these articles was the United States, specifically referencing presidential and vice-presidential candidates in the current election cycle.
Social Media Engagement
The operation used AI to generate short comments in English and Spanish for X and Instagram. The personas adopted varied roles to appear authentic:
- On X: Accounts posed as partisans for both main candidates in the U.S. presidential election.
- On Instagram: An account posed as a supporter of Scottish independence.
To support these activities, the accounts used ChatGPT for basic research, including seeking tips on how to improve social media engagement. In one instance, a user attempted to generate an article claiming that Microsoft's exposure of the operation proved Iranian cyber operations were more impactful than those of Russia or China, though OpenAI did not find evidence that this specific text was posted online.
Content Themes and Strategy
The generated content focused on highly polarized geopolitical and social issues to drive influence. Key themes included:
- Global Conflicts: The conflict in Gaza and Israel's presence at the Olympic Games.
- U.S. Politics: The U.S. presidential election and the rights of Latinx communities in the U.S.
- International Affairs: Scottish independence and politics in Venezuela.
To increase the perceived authenticity of the accounts and attempt to build a following, the actors also generated and posted generic content regarding beauty and fashion. OpenAI's DALL·E 3 classifier determined that the images used in these beauty posts were not generated by OpenAI services.
Impact and Reach
Despite the scale of content production, the operation did not achieve meaningful audience engagement. The majority of the social media posts garnered low to no engagement before the associated accounts were suspended, and there was no evidence that the long-form articles were widely shared on social media.