Leaked Files Detail Russia's 200,000-Page ChatGPT Poisoning Campaign

Russia's influence operators stopped trying to win your feed. They're building an alternative internet designed to poison what ChatGPT, Google, and every retrieval-augmented AI system treats as ground truth — and leaked Kremlin contractor files show the factory is already running.

The scale: Leaked documents from Russia's Social Design Agency — detailed in a June 23 Bloomberg investigation — describe Project 2026: eight active initiatives to flood search engines and large language models with Kremlin-aligned reference pages, fake think tanks, and cross-linked disinformation dressed up as independent research.

What's new

• Bloomberg reviewed 73 leaked SDA files spanning May 2023 to April 2026. Russian operators internally branded the AI-and-search push Project 2026.
• A Germany-focused "self-filling knowledge base" already contains more than 200,000 webpages, per OCCRP and Armenia's Fact Investigation Platform, which first reported portions of the same leak. Operators planned to edit 100 articles monthly for search visibility and aimed to "train" six AI platforms per month on the modified content.
• A phony think tank, the World Center for Strategic Studies, registered March 30, 2026, republishes reputable research with altered conclusions — including a rewrite of a French Institute of International Relations report that injected claims of a deepening European political crisis, per Bloomberg and UNITED24.

The mechanism

• Kateryna Sedova, an Atlantic Council senior fellow and former State Department official, told UNITED24: "Their approach is to try to break search engines by flooding the zone with content that cross-references their content or their narratives. This will be their indirect way of breaking into popular chatbots and search engines."
• FIP.am, reviewing the SDA materials, stated the explicit operational goal: make "artificial intelligences like ChatGPT tell what is beneficial to the Kremlin when asked questions about politicians."
• In Armenia, operators built Wikipedia-style clones — including sites Bloomberg identified as spyurk[.]cyou, sevan[.]info, and khachkar[.]info — and planned more than 50,000 SEO-optimized mirror pages ahead of parliamentary elections.
• France's 2026 plan calls for AI-generated video factories: 540 videos across six social networks, per OCCRP and FIP.am.

Not their first rodeo

• The SDA is the Kremlin-linked firm behind the Doppelgänger campaign. The U.S., U.K., and EU have sanctioned it; an FBI affidavit cited in OCCRP reporting described the agency as under the Russian presidential administration's "direction and control."
• The Atlantic Council's Digital Forensic Research Lab warned last year that Kremlin-linked Pravda fake-news networks were already rewriting Wikipedia and surfacing in AI chatbot answers across dozens of countries.
• Project 2026 scales that playbook from content farms to citation infrastructure — the pages models retrieve, not the posts users scroll past.

Why it matters now

• Chatbots are becoming news intermediaries. Reuters Institute data shows 10% of users now get news from AI weekly while only 4% click through to original publishers.
• Western intelligence chiefs warned June 22 that frontier AI transforms cyber offense and defense in months, not years — the same week OpenAI shipped GPT-5.5-Cyber to vetted defenders. Moscow is attacking a different seam: not hacking the model, but hacking what it reads.
• James Pamment, who directs Sweden's Psychological Defence Research Institute and has studied the SDA extensively, called the broader leak evidence of "a pattern of reckless escalation" in OCCRP's reporting — operations designed to inflame migration and religious tensions across NATO countries.

The gap

• Labs optimize for fluency and speed, not source provenance. Retrieval layers trust page rank, cross-link density, and apparent institutional formatting — exactly what a 200,000-page SEO factory exploits.
• No major chatbot provider has published a credible standard for detecting state-sponsored citation poisoning at web scale.

Convina's view: Disinformation's sharpest operators figured out that winning the feed was optional once you could win the bibliography. Project 2026 is not a social media stunt — it is industrial SEO aimed at the trust layer every AI answer engine depends on. Until provenance becomes a product requirement, not a press-release talking point, chatbots will keep laundering Kremlin fiction as analysis — and users will never know they left the chat window.