This is a research prototype. The data and analyses are preliminary and not yet validated — we'd welcome your .
A Moscow-based disinformation network known as 'Pravda' or 'Portal Kombat' is systematically flooding the internet with millions of automated, pro-Kremlin articles. This campaign, termed 'LLM grooming,' is designed to infiltrate the training data of major AI chatbots. Audits by NewsGuard and other researchers confirm that leading AI models frequently repeat these false narratives, effectively laundering Russian propaganda through AI systems.
A purported Moscow-based disinformation network, Pravda, allegedly infiltrated AI models by flooding the internet with pro-Kremlin falsehoods. A NewsGuard audit found that 10 major AI chatbots repeated these narratives 33% of the time, citing Pravda sources as legitimate. The tactic, called "LLM grooming," manipulates AI training data to embed Russian propaganda. Pravda is part of Portal Kombat, a larger Russian disinformation network identified by VIGINUM in February 2024, but in operation since February 2022.
Domain classification, causal taxonomy, severity scores, and national security assessments were LLM-classified and may contain errors.
AI systems that inadvertently generate or spread incorrect or deceptive information, which can lead to inaccurate beliefs in users and undermine their autonomy. Humans that make decisions based on false beliefs can experience physical, emotional or material harms
Also classified as
The Pravda network uses automated systems to conduct a large-scale disinformation campaign aimed at manipulating public opinion and AI models.
The chatbots fail to perform reliably under data poisoning conditions, demonstrating a lack of robustness in filtering out unreliable sources.
The incident involves major Western AI chatbots inadvertently repeating and spreading fabricated Russian propaganda narratives to users due to data poisoning.
AI system
Due to a decision or action made by an AI system
Unintentional
Due to an unexpected outcome from pursuing a goal
Post-deployment
Occurring after the AI model has been trained and deployed
The risk of chatbots outputting Russian propaganda occurs post-deployment due to the AI systems unintentionally absorbing and repeating poisoned web data during retrieval or generation.
Added categories peak at 4: Severe (Epistemic Harm, Indirect)
A Moscow-based disinformation network executed a massive 'LLM grooming' campaign, systematically poisoning the data environment of major Western AI chatbots. By publishing millions of automated articles, the network successfully caused leading AI models to repeat pro-Kremlin propaganda 33.5% of the time, demonstrating a significant evolution in state-sponsored information warfare.
The AI systems involved are large language model chatbots (such as ChatGPT, Gemini, and Copilot) acting as conversational assistants that provide information and suggestions to users, requiring human oversight and query inputs.
Limited Risk: The systems involved are chatbots and generative AI tools, which fall under Risk Level 3 due to transparency obligations to ensure users are aware they are interacting with an AI and that the content is AI-generated.
The primary function of the affected AI systems is acting as conversational chatbots and question-answering engines for users seeking news and information.
The reports provide detailed, consistent, and well-documented findings from multiple independent organizations (NewsGuard, American Sunlight Project, Viginum, EDMO) regarding the Pravda network's operations, scale (3.6 million articles), and the specific audit results showing a 33% replication rate across 10 major chatbots.
The incident could be viewed purely as a human-driven cyber/influence operation rather than an AI safety failure, though the deliberate targeting of LLM training data ('LLM grooming') makes it a distinct AI safety and robustness issue.
Technology
Vulnerability to 'LLM Grooming'
Models process massive token volumes, shifting probabilities toward falsehoods.
Inability to Verify Sources
Chatbots fail to distinguish reliable sources from laundered disinformation.
Dependence on Search Engines
Chatbots scrape real-time web data indexed by compromised search results.
Data Inputs
Data Poisoning via Automated Sites
Pravda network flooded web crawlers with 3.6 million articles in 2024.
SEO Manipulation of Training Data
Disinformation actors used SEO strategies to boost visibility for web crawlers.
Use of Unvetted Web Scraping
AI developers train models on unvetted, scraped public internet data.
Human Factors
User Overreliance on Chatbots
Users trust persuasive chatbot outputs without cross-checking other sources.
Deliberate Malign Prompting
Leading or malign prompts manipulate chatbots into bypassing safety guardrails.
Malign Intent of Propagandists
Propagandist Dougan intentionally engineered narratives to manipulate AI.
Process and Methods
Whack-a-Mole Domain Blocking
Blocking domains fails as the network continuously registers new subdomains.
Inadequate Training Safeguards
Companies lack robust technical mechanisms to vet data before model training.
Lack of Source Discrimination
Chatbot source citations treat reliable and unreliable sources equally.
Management
Prioritizing Speed Over Safety
AI companies deploy models rapidly without fully vetting training datasets.
Underinvestment in Data Vetting
Companies put insufficient energy into filtering poorly vetted scraped web data.
Regulatory Environment
Unregulated Chatbot Ecosystem
Lack of regulatory oversight allows chatbots to operate without safety rules.
Rollback of Cyber Defenses
US paused cyber operations against Russia, weakening foreign threat mitigation.
AI system involved
Identified by automated extraction; can differ from the alleged developers above.
