OpenAI Dismantles 5 Covert Influence Ops, Admits AI Didn't Help Much

OpenAI just dropped a transparency report revealing they’ve disrupted five covert influence operations in the last three months, and the news isn’t what you might expect. The company, which published its findings to promote cross-industry sharing, concluded that these propaganda mills didn’t actually get much of a boost from their models. Using the Brookings Breakout Scale, none of the busted campaigns scored above a 2 out of 6, meaning they were active on multiple platforms but failed to break into any authentic communities. In the end, designing models with safety in mind often blocked the threat actors from generating the content they wanted in the first place.

The operations spanned the globe and covered a grim bingo card of geopolitical hot spots. A previously unreported Russian group OpenAI calls “Bad Grammar” used its models to debug a Telegram bot and pump out political comments targeting Ukraine and the Baltic states. The known Russian network Doppelganger generated multi-language comments for X and 9GAG. China’s Spamouflage, an Iranian operation called IUVM, and a commercial Israeli company nicknamed “Zero Zeno” were all caught using the tools for tasks ranging from spinning up fake social media bios to translating articles and debugging database code. The content pushed narratives on Russia’s invasion, the Gaza conflict, Indian elections, and US and European politics.

Crucially, OpenAI is framing this as a defensive win not just for them, but for the ecosystem. They credit years of detailed threat reporting by distribution platforms and the open-source community for making their own investigations possible. The irony is palpable: the same AI tools that malicious actors tried to weaponize are simultaneously making OpenAI’s internal forensic teams more efficient at hunting them down. It’s a cat-and-mouse game where the cat is increasingly using a super-powered flashlight.

For all the hype around AI-generated deepfakes destroying democracy, this report suggests the current reality is far more pedestrian. These are largely low-effort, spam-level operations getting caught in the net. The real leverage for defenders might not be the AI models themselves, but the rapid intelligence-sharing they enable.