OpenAI has banned a cluster of accounts it says used its models to churn out English-language comments attacking Russia’s Anti-Corruption Foundation, the group founded by the late Alexei Navalny. The operation, which OpenAI dubbed “Corrupt Comment” in its October 2024 influence operations report, generated generic criticisms questioning the FBK’s transparency and then posted them via fake accounts on X. A key operational blunder made detection straightforward: the English comments were often dropped as replies to posts that were originally in Russian, creating a jarring mismatch that screamed automation. OpenAI noted it saw no evidence the operator used its models to actually read or analyze social media content to craft relevant responses—this was a blunt “theme and variations” spam technique, not a sophisticated conversational AI deployment.
The sockpuppet accounts themselves were flimsy. Most were created in a single burst in December 2023, had zero followers, and received zero replies. Profile pictures were either generic scenery scraped from around the web or faces generated with an older era of AI, specifically generative adversarial networks (GANs), a technique whose telltale artifacts were widely reported on as far back as 2019. Using such dated image generation to support a modern large language model operation reveals a sloppy operational security posture. If the intent was to drown out genuine discourse, it failed completely—OpenAI assessed that English-language replies from this network were usually outnumbered by organic Russian-language replies from unrelated third parties.
The tactical specifics here point to a low-budget, low-impact influence attempt rather than a well-resourced state program. OpenAI rated the activity a 1 on the Breakout Scale, the lowest tier for influence operations, meaning it generated content but achieved no meaningful amplification. This stands in stark contrast to more persistent covert campaigns that OpenAI and other firms have disrupted, where operators invested heavily in building believable personas and nurturing audiences over time. The laziness of the approach—zero-follower accounts, language mismatches, GAN-generated faces—suggests either a rushed proof-of-concept or an operator with very limited resources. As Meta and Google have also documented in their quarterly adversarial threat reports, the barrier to entry for AI-enabled influence ops is collapsing, even if the execution often remains amateurish.
The Navalny connection gives this particular case a sharp political edge. Since the opposition leader’s death in a Russian penal colony in February 2024, the FBK has remained a high-priority target for pro-Kremlin information operations both online and off. What distinguishes this operation is the toolchain: a US-made foundation model deployed against a US-designated anti-corruption group. OpenAI’s willingness to name the specific target and provide granular operational details continues its recent shift toward transparency in influence operations reporting, a practice that was once limited to companies like Meta. For cybersecurity teams tracking these threats, the report offers a useful signature: GAN-generated avatars combined with GPT-generated text in language-mismatched replies is a pattern worth hunting for.
