AI News AI资讯 2h ago Updated 2h ago 更新于 2小时前 49

Russia used ChatGPT to run a covert influence campaign pushing pro-Kremlin narratives across the West 俄罗斯利用ChatGPT开展秘密影响力运动,向西方向传播亲克里姆林宫叙事

OpenAI exposed a covert Russian influence campaign that used ChatGPT to generate pro-Kremlin social media content while deliberately masking Russian linguistic markers The campaign promoted the "International Burke Institute" (IBI), a fake think tank ostensibly based in Israel, which published a "sovereignty index" ranking Russia favorably against Western nations 34 of 36 IBI articles were plagiarized from legitimate sources with fabricated author credentials, including misattributed papers from 俄罗斯利用ChatGPT生成社交媒体内容,通过VPN隐藏来源并指示模型消除俄语特征,开展亲克里姆林宫叙事活动 OpenAI封禁相关账号并揭露虚假智库"国际伯克研究所"(IBI),其34/36篇文章抄袭自其他来源且作者身份造假 活动跨X、LinkedIn、Facebook、Substack和Telegram等多平台传播,当前影响力有限但基础设施具备可扩展性 这是OpenAI继"Bad Grammar"和"Operation Helgoland Bite"后第三次揭露俄罗斯利用ChatGPT进行影响力操作 实际AI驱动影响力操作规模可能更大,因中国AI提供商和开源模型更易被滥用

72
Hot 热度
68
Quality 质量
70
Impact 影响力

Analysis 深度分析

TL;DR

  • OpenAI exposed a covert Russian influence campaign that used ChatGPT to generate pro-Kremlin social media content while deliberately masking Russian linguistic markers
  • The campaign promoted the "International Burke Institute" (IBI), a fake think tank ostensibly based in Israel, which published a "sovereignty index" ranking Russia favorably against Western nations
  • 34 of 36 IBI articles were plagiarized from legitimate sources with fabricated author credentials, including misattributed papers from Cambridge University Press and the Migration Policy Institute
  • The operation spread across X, LinkedIn, Facebook, Substack, and Telegram with low individual post reach but an elaborate, scalable infrastructure capable of targeting multiple countries
  • This follows a pattern of prior Russian AI-powered disinformation campaigns, including "Bad Grammar" (2024) and "Operation Helgoland Bite" (2025), with the true scale likely far larger given access to Chinese AI providers and open-weight models

Why It Matters

This case illustrates how generative AI lowers the barrier for state-sponsored influence operations, enabling actors to produce large volumes of persuasive, linguistically polished content at scale while evading detection. For AI practitioners and platform operators, it underscores the urgent need for robust abuse detection systems that can identify coordinated inauthentic behavior powered by LLMs. The incident also highlights a growing geopolitical threat vector where AI tools become weapons of information warfare, with implications for election integrity and public discourse in democratic nations.

Technical Details

  • The operators accessed ChatGPT from Russia via VPNs and used explicit prompts instructing the model to eliminate Russian linguistic markers, demonstrating deliberate adversarial prompting to evade geolocation and stylistic detection
  • The "International Burke Institute" served as a fabricated authority figure, with plagiarized content falsely attributed to academics at institutions like the University of Nottingham and an Australian food chemistry professor, creating an illusion of expert credibility
  • Content was distributed across five major platforms (X, LinkedIn, Facebook, Substack, Telegram), with a German-language Telegram channel ("Lahme Ente") targeting audiences in Germany, while a second operator managed branding for approximately a dozen country-specific Telegram channels
  • OpenAI rated the campaign at category 3 out of 6 on the Brookings Breakout Scale, indicating multi-platform spread with early organic reach, while linked Telegram channels each accumulated 10,000–20,000 followers
  • The campaign's infrastructure included logo design, cross-platform content coordination, and Russian-language activity summaries, suggesting a structured operational workflow rather than ad hoc posting

Industry Insight

  • AI platforms must invest in multi-layered detection systems combining linguistic forensics, behavioral pattern analysis, and cross-platform correlation to identify coordinated AI-assisted influence operations before they achieve breakout scale
  • The availability of open-weight models and Chinese AI providers means containment on a single platform is insufficient; industry-wide threat intelligence sharing and standardized abuse-detection protocols are essential to address the distributed nature of the threat
  • Organizations and governments should treat AI-generated disinformation as a persistent operational risk, developing rapid response capabilities and media literacy initiatives to help the public identify fabricated expert sources and plagiarized content masquerading as original analysis

TL;DR

  • 俄罗斯利用ChatGPT生成社交媒体内容,通过VPN隐藏来源并指示模型消除俄语特征,开展亲克里姆林宫叙事活动
  • OpenAI封禁相关账号并揭露虚假智库"国际伯克研究所"(IBI),其34/36篇文章抄袭自其他来源且作者身份造假
  • 活动跨X、LinkedIn、Facebook、Substack和Telegram等多平台传播,当前影响力有限但基础设施具备可扩展性
  • 这是OpenAI继"Bad Grammar"和"Operation Helgoland Bite"后第三次揭露俄罗斯利用ChatGPT进行影响力操作
  • 实际AI驱动影响力操作规模可能更大,因中国AI提供商和开源模型更易被滥用

为什么值得看

本文揭示了大语言模型被用于地缘政治信息战的典型案例,对AI安全研究和平台治理具有重要参考价值。OpenAI的应对策略和评估框架为行业提供了可借鉴的威胁响应模式,同时凸显了AI时代信息操纵的新挑战。

技术解析

  • 隐蔽操作技术:俄罗斯操作员通过VPN访问ChatGPT,并在提示词中明确要求模型"隐藏任何指向俄罗斯起源的语言线索",体现了对AI模型输出可控性的针对性利用
  • 虚假内容生产:创建的"国际伯克研究所"(IBI)智库发布"主权指数"排名,其中34/36篇专家关联文章(2025年9月-2026年5月)抄袭自其他来源,包括将剑桥大学出版社文章错误归因于诺丁汉大学教授、将移民政策研究所文章归因于澳大利亚食品化学教授
  • 多平台分发架构:活动覆盖X、LinkedIn、Facebook、Substack和Telegram,其中德语Telegram频道"Lahme Ente"(跛鸭)针对乌克兰、欧盟和德国政府,另一操作员为12个聚焦德国、美国、法国、波兰和土耳其的Telegram频道设计标识
  • 影响力评估:采用Brookings Breakout Scale评级为3/6级,表明跨多平台传播且开始触及真实用户,但单篇帖子浏览量极低,IBI官方账号订阅者少,关联Telegram频道各有1-2万粉丝
  • 历史案例对比:继2024年6月"Bad Grammar"(针对俄罗斯、乌克兰和波罗的海国家Telegram评论)和2025年"Operation Helgoland Bite"(德国联邦选举前生成德语内容攻击美国/NATO并推广AfD党)后,这是第三次类似行动

行业启示

  • AI治理紧迫性:大模型正被系统性用于地缘政治信息战,平台需建立更严格的身份验证、内容溯源和异常模式检测机制,防止技术被武器化
  • 跨国协作需求:影响力操作依赖开源模型和免费AI服务,单一平台应对措施有限,需加强国际监管合作与威胁情报共享,尤其关注中国AI提供商和开源模型的监管漏洞
  • 技术防御方向:应发展AI生成内容检测、提示词注入防御和跨平台行为分析技术,同时建立"AI影响力操作"的标准化评估框架,提升行业整体防御能力

Disclaimer: The above content is generated by AI and is for reference only. 免责声明:以上内容由 AI 生成,仅供参考。

GPT GPT Security 安全 Policy 政策 Ethics 伦理