OpenAI Disrupts STORM-2035 Recidivist Influence Operation

OpenAI has banned accounts associated with STORM-2035, a recidivist influence operation likely linked to Iran. The operation used ChatGPT to generate short, multi-lingual comments for distribution via inauthentic accounts on X (formerly Twitter) to influence public opinion in target countries.

Operation Mechanics and Actor Behavior

STORM-2035 utilized ChatGPT to generate small batches of tweets—typically a dozen at a time—based on prompts written in Persian. The actor requested outputs in either English or Spanish, which were then posted by accounts posing as residents of the United States, United Kingdom, Ireland, and Venezuela.

Key characteristics of the inauthentic accounts included:

  • Account Activity: The accounts exclusively posted content generated by this operation.
  • Follower Metrics: Each account followed significantly more users than it had followers.
  • Visual Identity: Many profiles used images of young women with faces obscured by phones or other objects, with some images appearing to be copied from Pinterest.

Compared to the original disruption of this actor last August, the recidivist operators demonstrated a shift in strategy by attempting to isolate workstreams. While the original activity spanned X, Instagram, and various websites, this specific recidivist activity was limited exclusively to X.

Content Themes and Target Demographics

The operation generated content across a range of geopolitical and social topics, tailored by language to specific target regions:

  • Latino Rights: Spanish-language comments targeting the US, focusing on inclusion and criticizing the politics of exclusion and specific political figures.
  • Iran's Military Prowess: Spanish-language comments portraying Iranian military power as the catalyst for forcing the US to negotiate.
  • Venezuela and Cuba: Spanish-language comments contrasting Cuban medical exports with US military exports and discussing President Nicolás Maduro.
  • Scottish Independence and UK Criticism: English-language comments targeting the UK.
  • Irish Reunification: English-language comments targeting Ireland.
  • Palestinian Rights: Comments generated in both English and Spanish.

OpenAI noted one instance of potential operator error where an account typically focused on Scotland (English) switched to posting in Spanish regarding Palestinian rights.

Impact Assessment

The influence operation had minimal reach and engagement. Typical tweets generated by STORM-2035 recorded between 150 and 350 views, with zero likes, shares, or comments.

OpenAI found no evidence that real users amplified the content. Using the Brookings Institution's Breakout Scale for measuring influence operations, OpenAI categorized this activity as Category 1: activity limited to one platform with little evidence that real people picked up or widely shared the content.

Sources