이미지: METAL LAB 생성
Summary
- OpenAI has blocked a cluster of ChatGPT accounts that connected from Russia through VPNs. The accounts were promoting a fake think tank called the International Burke Institute (IBI) that claimed to be based in Israel.
- Of 36 expert-bylined articles posted on the IBI site, 34 turned out to be copied from other sources, and the site also spread a "sovereignty index" that portrayed Russia favorably.
- The content spread across X, LinkedIn, Facebook, Substack, and Telegram, with affiliated Telegram channels each gathering 10,000 to 20,000 subscribers. OpenAI said what stood out wasn't the reach but how sophisticated the infrastructure was.
- 적발 주체·시점
- 오픈AI, 2026년 8월 25일 발표
- 차단 대상
- 러시아發 챗GPT 계정 클러스터, VPN으로 접속
- 가짜 싱크탱크
- 국제버크연구소(IBI), 2025년 2월 도메인 등록, 이스라엘 소재 자칭
- 표절 비율
- 2025년 9월~2026년 5월 전문가 연계 기사 36건 중 34건 타 출처 복제
- 확산 플랫폼
- X, 링크드인, 페이스북, 서브스택, 텔레그램
- 텔레그램 채널 규모
- 연계 채널당 구독자 1만~2만 명
- 영향력 평가
- 브루킹스 브레이크아웃 스케일 6단계 중 3단계
- 과거 유사 사례
- 배드 그래머(2024년 6월), 오퍼레이션 헬골란트 바이트(2025년)
OpenAI Blocks a Cluster of ChatGPT Accounts Tied to Russia
On August 25, OpenAI announced it had taken down a cluster of ChatGPT accounts that appeared to originate from Russia. The accounts had been using ChatGPT to promote a fake think tank calling itself the International Burke Institute (IBI), which claimed to be headquartered in Israel, and to spread a "sovereignty index" that cast Russia in a favorable light while disparaging the West.
OpenAI said the investigation started with tracking AI-generated social media posts and led to a much larger influence operation than expected. This isn't the first time OpenAI has disclosed a Russia-linked campaign. In June 2024, the company exposed a network called "Bad Grammar" that generated political comments about Russia, Ukraine, and the Baltic states, and in 2025 it flagged "Operation Helgoland Bite," which spread German-language content ahead of Germany's federal election criticizing the US and NATO while backing the AfD. In the report OpenAI published, the company said "the sophisticated setup distinguishes this operation from other Russia-linked campaigns."
Sophisticated Camouflage Behind a Fake Think Tank
IBI's website registered its domain and began operating in February 2025. The site listed prominent scholars such as Francis Fukuyama and Noam Chomsky as experts, but most of its actual content turned out to be lifted from elsewhere. When OpenAI sampled 36 expert-bylined articles posted between September 2025 and May 2026, 34 of them turned out to be copied verbatim from other sources. In one case, an original piece on the China-Pakistan Economic Corridor published by Cambridge University Press was misattributed to a University of Nottingham professor. In another, an article on migration governance written by the Migration Policy Institute was passed off as the work of an Australian professor whose actual specialty was food chemistry.
The IBI site interspersed this plagiarized content with "sovereignty index" reports on individual countries. It targeted nations that had criticized Russia's invasion of Ukraine, such as France, Germany, the EU, and the United States, taking a blunt editorial line — comparing Macron's government to a custodian for Rothschild banking interests and criticizing Germany's coalition government for its military dependence on the US.
ChatGPT Was Used Both to Generate Content and to Mask Identity
According to OpenAI, the operators worked around the fact that ChatGPT access is blocked in Russia by connecting through VPNs. They typed prompts in Russian to generate English-language social media comments and instructed ChatGPT to erase any traces suggesting the writer was a Russian speaker. The generated posts were distributed not only through accounts branded with IBI's name and logo but also through accounts posing as ordinary users, spreading across X, LinkedIn, Facebook, Substack, and Telegram. On Substack, ChatGPT was even used to draft comments on real users' posts urging them to subscribe to IBI channels.
One operator posted content to a German-language Telegram channel called "Lahme Ente" (lame duck) that criticized Ukraine, the EU, and the German government while calling for improved relations with Russia. Another operator used ChatGPT to generate logos for a dozen or so Telegram channels targeting Germany, the US, France, Poland, and Turkey, and repeatedly asked the model to summarize channel activity in Russian. The bio for one of these channels, which impersonated a US media outlet, contained several instances of unnatural English phrasing.
Small Reach, But Sophisticated Infrastructure
OpenAI said the view counts on individual posts and the subscriber numbers for IBI's official accounts were both fairly low. Still, the affiliated Telegram channels each managed to attract between 10,000 and 20,000 subscribers. Using the Brookings Breakout Scale, a six-tier framework for measuring influence operations, OpenAI ranked this campaign in the lower half of tier three — meaning it had spread across multiple platforms and shown early signs of reaching some authentic audiences.
OpenAI weighted the significance of the operation less on its actual reach and more on the fact that it had built infrastructure designed to look like a genuine institution complete with experts, academic-style content, and its own index. ChatGPT was only used to generate posts promoting that infrastructure, but OpenAI's assessment is that those posts were part of a larger effort to fabricate a convincing sense of authority.
Russia-Linked Operations OpenAI Has Disrupted
| Operation | Disclosed | Main Activity | Target Platforms |
|---|---|---|---|
| Bad Grammar | June 2024 | Generated political comments about Russia, Ukraine, and the Baltic states | Telegram |
| Operation Helgoland Bite | 2025 (ahead of German election) | Generated German-language content criticizing the US/NATO and supporting AfD | Online media |
| IBI Influence Campaign | August 2026 | Promoted a fake think tank and spread "sovereignty index" content | X, LinkedIn, Facebook, Substack, Telegram |
Editor's Take
This is the third Russia-linked operation OpenAI has disclosed, following Bad Grammar in June 2024 and Operation Helgoland Bite in 2025. Lined up together, the three cases show a clear progression. What started as posting political comments on Telegram moved on to backing a specific party ahead of Germany's election, and now has evolved into building infrastructure that mimics an entire academic institution. That trajectory suggests AI's role in influence operations is shifting from a simple content-generation tool to something used to manufacture credibility itself.
There are two practical takeaways here. First, the thing worth scrutinizing isn't just AI-generated content itself but the credibility of the "source" that content points back to. A site like IBI, dressed up with the names of prominent scholars and academic-sounding papers, doesn't look fake at a glance. Companies and institutions citing outside reports probably need to get in the habit of tracing sources back one more step. Second, as OpenAI itself acknowledged, this operation was caught almost by accident, while the company was tracking activity happening on top of ChatGPT. If OpenAI blocks access to its own platform, the same operation could simply migrate to other open-weight models or Chinese-made models and continue outside anyone's line of sight — meaning the actual scale of this kind of activity is likely far larger than what's been exposed so far.
It wouldn't be surprising to see similar disclosure reports from other major tech companies in the coming weeks. If each platform keeps disclosing only the activity it can see on its own turf, influence operators will have even more incentive to shift toward open-source models where tracking is looser.
Comments