OpenAI Discloses Shutdown of Covert Russian Influence Operation Leveraging ChatGPT to Fabricate a Fake International Think Tank, Undermining the Future of Disinformation
In a decisive and revealing move, OpenAI has formally banned a cluster of ChatGPT accounts that operated from within Russia, dismantling a covert and meticulously orchestrated influence campaign intended to advance narratives favorable to the Kremlin. The company revealed in a comprehensive threat report that the operators, working in the Russian language, bypassed OpenAI’s geographic restrictions—which explicitly prohibit access to its models from within Russia—by routing their connections through virtual private networks (VPNs). This initial layer of circumvention allowed them to harness the generative power of ChatGPT to craft a vast, polished body of social media content intended for international audiences, primarily in English. The volumes of automated text were subsequently distributed across major global platforms such as Substack, Telegram, X (formerly Twitter), Facebook, and LinkedIn. Rather than a momentary engagement with troll networks, open looking at this operation, OpenAI identified a remarkably structured and patient strategy, one that did not yearn for instantaneous viral reach but instead built a layered, brutal façade of legitimacy for fabricated political analyses and strategically planted pro-Russian narratives, a concerning indicator that AI’s semantic qualities have entered the queue of ammunition for state-driven espionage.
At the heart of this deceptive campaign was the fabrication of a complete, state-sponsored think tank called the International Burke Institute (IBI). The institute, its website, and its public profile were crafted to superficially emulate the characteristics of a legitimate academic, geopolitical expert community. According to the description, the website claimed a physical street address in Israel—a geographic detail presumably chosen to give the enterprise a false sense of geopolitical neutrality and non-affiliation with Moscow. Further entrenching the illusion, the site falsely (listed) and prominently featured the names of eminent scholars, including the renowned political scientist Francis Fukuyama and critical theorist Noam Chomsky, as purported contributors or members. Deep-diving into the content, OpenAI conducted a meticulous review of 36 articles between September 2025 and May 2026. The findings revealed the extent of institutional sycophancy: 34 of these articles were outright plagiarized, lifted from other online sources, and often furnished with incorrect or fabricated attribution, a practice that deliberately launders the citation system that establishes academic veracity. The core intellectual utility of the institute, however, came from a bespoke, freshly-invented quantitative measure its founders called the “sovereignty index,” a ranking metric calculating the geopolitical autonomy of various countries. Predictably, the index was designed to immediately calendar of Russia and explain it favorably, while simultaneously castigating Western nations, particularly the United States, Germany, France, and the broadswept European Union, with reports that veered into strident, polemics rather than objective analysis, setting up a completely tailored, adversarial transparent taxonomy of global power status.
The detailed content production assets extended beyond the fabricated papers. OpenAI’s investigations unveiled a specific composition of operators who split specialized tasks to double the operation’s narrative scope. One operator was responsible for generating German-language posts for a dedicated Telegram channel ominously named “Lahme Ente,” translating to “Lame Duck.” This channel becomes a hate-platform targeting the Ukrainian government, the European Union leadership, and NATO and German political classes, while simultaneously propagandizing for the strengthening of economic and political ties with Russia. The military nuance of this is not a random stirring of the pot; it is a tailored psychological operation aimed at inflaming division within European coalition and appealing to fringe domestic beings. A second operator in the group executed a comprehensive visual branding endeavor, creating logos and distinct aesthetic identities for roughly a dozen separate Telegram channels that would gain traction in Germany, the United States, France, Poland, and Turkey. After establishing these respective presence channels for content—each seemingly a unique grassroots initiative—the operator then requested Russian-language reports on their activity, synthesize data on engagement and post-performance to facilitate operational nuance, demonstrate a centralized command structure with kite, even as the local officials are presented as decentralized, organic civil-society impulses. This operational blue fingerprint is a hallmark of classic intelligence cycles, but now turbocharged by LLM to automate creativity based on a continuous feedback loop, allowing spec campaigns to mutate and recalibrate within hours based on regional developments.
A notable entrance factor in this episode is the astonishing dissonance between the claimed technical sophistication and the direct, quantitative reach it achieved. OpenAI asserts that most of the posts on these channels received minuscule view counts—and the flagship official accounts for the connected institutes maintained only a handful of subscribers. The Telegram channels attracted between 10,000 and 20,000 followers each, which is certainly a modest size relative to hulking country-wide ones—Italy has reached millions. Yet OpenAI explicitly rejects the empty measure of reach as the primary metric of concern. The underlying threat, it argues, lies in the infrastructure that this template creates—a node for later, infinitely scalable operations. By injecting branded, polished digital corpus into these channels, the operators solidifying a preature ground in which, after a certain threshold, the same content can be amplified, lifted into major global loops, and gain traction through searching platforms. The realistic “News” from a fake thinktank that plagiarized authors and has a statistical index makes the lies difficult to overcome; presenters can point to the “institute” as a genuine intellectual source to news outlets that routinely use academics as fixers for injury content. This enterprise ethos indicates that the current challenge is no longer creating boundary bots but creating “AI-that-vives-tanks,” an embalmment that exist in the public consciousness as a rationale of truth, allowing for succession: future operations to amplify it, enhancing their capabilities from a baseline, instead of building from scratch each time.
From OpenAI’s perspective and the global infosec community, this marks a turning point in how AI models are now appropriating low “cost of policies” to influence operators. Historically, building a fake think-tank required immense hire of consultants, cybermarket price drivers, maintaining webhost, hiring local webmasters for languages like German and Polish, including OCR and PDF document production. With ChatGPT, these Russian individuals did the work of an entire embassy’s PSYOP division in a fraction of time, maintaining and customizing output to precisely suit differing regional cultures and sensibilities, and directing the model to remove any stylistic traces of machine generation that may betray an origination. By scanning on OpenAI, account bans are a celebration but remain limited in utility. The ban that took place can be refulated within weeks of purchase: new VPN progies, new TempMail, and ChatGPT Official documentation for savvy users eventually in scripts to fine-tune the standard model. But it cuts into its spending and throttle data usages. However, the nets and pipelines are captured and dated. Most concerningly, this operation signals a pivot away from the soapbox use of cheap synthetic social media botches. Threats actors are pulling their investment out of coordinated spam to forge synthetic authentic expression, one that deceives not just social media algorithms but human instincts of peer computing.
Ultimately, the action is a stark reminder that intelligence operations are no longer confined to reality—they now occurs inside the code layer, creating shadows across our public knowledge of the world. The takedown by OpenAI stops the mute flow of harmful outputs but leaves a gaping evidence trail of the modus operandi for advanced-state militaries, showing them that the bluff of vector influence now takes place through subjective aggregates. The usage of VPNs and multilingual content generation signals that Russian actors retain high adaptive creativity, relying on the prediction sandbox of the West itself to build a sanctuary. While the immediate audience numbers are negligible, the persistence and replication of that model illuminate a broader tectonic shift: the inevitability of AI systems being politically weaponized. For platforms, intelligence agencies, and media researchers, the primary means of countering this orthogonal threat is not simply re-sorting technical blocks; it demands the eruption of typographic interconnection, authentication of authorship with scientific institutions, and the cultivation of critical media literacy accessible to regular readers to parse that a “printing press” de la environment where fake thinktank formerly praised is no longer a future dissonance, but occurred slightly sooner than expected. The micro-scale phenomenon (10-20k followers) is now a laboratory—a petri dish where this synthetic, platform shepherded sovereignty index and plagiarized academic structure can perfect its baiting techniques—propagating into a landscape that the open public is quickly losing the tools to distinguish the genuine from the fabricative, as its prime manipulation may operate legally.