150 Accounts, One Playbook: OpenAI Uncovers the Multilingual 'A2Z' Influence Op

Hi, it's Shiichan! Today I want to walk you through an influence operation that ran across several languages at once.

OpenAI News openai.com

What was announced?

OpenAI's News page announced that it had banned a previously unreported cluster of accounts. This group used OpenAI's API to generate comments and post them on X and Facebook. A lot of the content praised Azerbaijan or pushed back on criticism of its human rights record, but the operation also covered a surprisingly wide range of other topics, which suggests it may have been commercially run. Because it praised Azerbaijan so often and tried out AI at multiple stages of its operation, OpenAI named it "A2Z."

Why it matters

Like the Russian-language Telegram operation "Bad Grammar" reported back in May, this network used AI to manage a large number of fake social-media personas at once, generating bios, analyzing posts and comments, and drafting and proofreading replies in many languages. Before AI, an operation at this scale would have needed a sizable team of human trolls. Now a much smaller team can run it, and that shift is exactly the risk this case illustrates.

What changes

OpenAI banned the accounts tied to this network. Because the operation leaned on AI at so many points in its pipeline, cutting off API access broke several capabilities at once. After OpenAI disrupted the activity in early June, the accounts it had identified stopped posting through the election periods in the EU, UK, and France, with some only resuming in late August. On the Breakout Scale, OpenAI placed it at the high end of Category 2, close to Category 3.

Dive Deep

OpenAI's findings get quite specific about A2Z's scale and tactics.

• Around 150 accounts were identified across X and Facebook, with more suspected
• Most posts were short comments, but the operation also generated stylized images resembling 1930s propaganda posters (though these typically drew just 0-5 engagements)
• Turkish-language accounts occasionally had short exchanges with real people
• The dominant theme was praise for Azerbaijan's government, from promoting unity among Central Asian countries to hosting a UN climate conference, while English- and Italian-language accounts defended Azerbaijan as an energy supplier to Europe
• Other threads covered AI, financial markets, environmental issues, cryptocurrency, and framed the UAE as a stable investment destination
• It also touched US and European politics with inconsistent leanings: one English account criticized Trump from a liberal angle, while French-language accounts criticized Macron and backed the National Rally party
• Among the roughly 150 sampled accounts, the largest following on X was just 222, and the highest reaction count on a Facebook post was 36, most got far less

With no consistent ideology and stances that shifted by region, the real goal here looks less like pushing one political message and more like generating broad, multilingual engagement for its own sake.

Wrap-up

• OpenAI banned A2Z, a network that posted multilingual comments mostly praising Azerbaijan
• Around 150 accounts spanned X and Facebook, with AI handling posts, image generation, and even replies to real users
• The Breakout Scale rating was the high end of Category 2, nearing Category 3, based on multi-platform activity and some engagement
• After the ban, the network's accounts fell silent through election periods in the EU, UK, and France
• A good example of how heavily AI-dependent operations can be disrupted all at once with a single takedown!