JA EN
shiichan

150-350 Views, Zero Likes: OpenAI's STORM-2035 Repeat Offender Flopped

OpenAI News ★★★ View Markdown
#openai#ai-safety#security

Hi, it's Shiichan! Today's story is about a repeat offender that came back... and still barely made a dent. Let me walk you through OpenAI's latest report on countering misuse of AI.

OpenAI News openai.com

What was announced?

OpenAI's News reported that it banned accounts tied to recidivist activity from STORM-2035, a likely Iran-linked influence operation. This is the same network OpenAI first disrupted around August 2024. This time, the operators prompted ChatGPT in Persian to generate short comments in English and Spanish, then posted them through a series of fake accounts on X (formerly Twitter).

The story so far

The original STORM-2035 activity spanned X, Instagram, and several websites, spreading both 700-900 word articles and short comments. This latest round of recidivism, by contrast, stuck to X only. According to OpenAI, the operators seem to have made a greater effort to keep their different workstreams separate this time around, suggesting they adjusted their approach after getting caught before.

What changes

OpenAI banned this new batch of fake accounts as well. The accounts posed as residents of the US, UK, Ireland, and Venezuela, and shared some telling patterns: each followed far more accounts than it had followers, and many used profile pictures of young women with their faces obscured, some apparently copied from elsewhere online, including Pinterest. Documenting these patterns could help spot similar networks going forward.

Dive Deep

Here's how the operation worked:

  • Prompted ChatGPT in Persian, asking for output in English or Spanish
  • Generated batches of about a dozen tweets at a time
  • Posted the resulting comments through accounts impersonating residents of the target countries

The topics and languages broke down by region:

  • Latino rights, Iran, Cuba, and Venezuela topics: in Spanish
  • Scottish independence and Irish reunification topics: in English
  • Palestinian rights topics: in both English and Spanish

Some posts framed Iran's military and diplomatic strength as having forced the US to the negotiating table.

The actual impact was underwhelming. Typical tweets recorded 150-350 views with zero likes, shares, or comments, and OpenAI found no evidence that real users amplified the content. Using the Breakout Scale, a framework for measuring the reach of influence operations, OpenAI rated this activity as Category 1: confined to a single platform, with little sign that real people picked it up.

Wrap-up

  • OpenAI banned accounts tied to recidivist activity from STORM-2035, the Iran-linked network it first disrupted in 2024
  • This round stuck to X only, using ChatGPT prompted in Persian to generate short English and Spanish comments
  • Fake accounts posed as residents of the US, UK, Ireland, and Venezuela, covering topics from Latino rights to Iran's diplomacy
  • Posts drew just 150-350 views each, with zero likes, shares, or comments
  • OpenAI rated the operation Category 1 on the Breakout Scale, meaning it barely spread at all

If you're interested in how AI providers catch and stop misuse, and how influence operations get measured, this report is worth a read!

Related posts