OpenAI busts “A2Z” network using its own API to pump pro-Azerbaijan propaganda

OpenAI has pulled the plug on a sprawling, commercially-run influence operation it’s calling “A2Z,” a network that used the company’s own models to generate content defending Azerbaijan’s government across X and Facebook. The takedown, detailed in OpenAI’s October 2024 threat report, is a case study in the ugly side of generative AI’s efficiency. The operation wasn’t just posting simple spam; it used AI to manage fake personas, craft bios, and even engage in real-time arguments with actual people in comment threads. The content ranged from praising Azerbaijan as a reliable European energy supplier to criticizing French President Macron, with the ideology shifting opportunistically depending on the target audience.

What makes A2Z notable isn’t its reach—the roughly 150 identified accounts were largely shouting into the void. A typical post on X garnered single-digit engagements, and the most followed account OpenAI found had a paltry 222 followers. Some accounts did manage to pull real people into short conversations, where the AI-generated replies were sometimes met with direct pushback. The network also produced stylized images mimicking 1930s posters, though OpenAI assesses these were designed for engagement, not deepfakes. The sheer volume of themes—from crypto and finance to the US election and Russia’s war in Ukraine—strongly suggests a for-hire operation rather than a single state-directed campaign.

This reliance on AI was also the operation’s fatal weakness. Because A2Z wove the technology through nearly every step of its content creation, OpenAI’s single disruption in early June effectively shattered the entire chain at once. The identified accounts went silent throughout the critical EU, UK, and French election periods. Some flickered back to life in late August, and a few posts then showed AI watermarks from other models, but the core operation was broken. It’s a stark demonstration that while AI lets bad actors scale cheaply, that very dependency creates a massive, centralized point of failure.

The A2Z case blows a hole in the assumption that AI-driven influence ops are always about reaching millions. The real story here is about accessibility: a level of multilingual, reactive trolling that once required a room full of people now appears viable as a low-cost service. The fact that the network touched everything from US partisan politics to UAE investment pitches suggests we’re looking at the early, messy prototypes of a disinformation-as-a-service industry. The low engagement numbers shouldn’t be comforting; they show the barrier to entry has collapsed, and it’s only the quality of the targeting, not the tech, that’s lagging behind.