A sprawl of near-empty websites that almost no human visits, built not to persuade people but to flood search results, pollute Wikipedia, and quietly seed the answers that AI chatbots give millions of users.
Most disinformation is designed to be read. Portal Kombat was not. When France’s government watchdog for foreign interference first mapped the network, what stood out was how little of it seemed aimed at a human audience at all. The sites carried almost no original writing, drew almost no visitors, and republished the same pro-Kremlin material over and over across hundreds of near-identical domains. It looked less like a propaganda outlet than like a machine for manufacturing the appearance of consensus, and that turned out to be exactly what it was.
What VIGINUM found
In February 2024, VIGINUM, the technical service inside France’s national defence secretariat, published a report naming the network Portal Kombat. It described a structured and coordinated ecosystem of pro-Russian sites, many built around the word “pravda,” that did not produce journalism. Instead they aggregated content from three sources: Russian state media, Russian officials’ social accounts, and Telegram channels. The network republished this material automatically, at high volume, in a growing list of languages, and it timed the launch of new country-specific domains to elections and to Russian strategic interests.
The infrastructure pointed back to Russia. Researchers linked the operation’s technical backbone to a Crimea-based IT company, and the network kept expanding. By late 2024 the ecosystem had grown into the hundreds of aggregator sites, with domains targeting more than eighty countries, launched in official and minority languages and especially active during election periods. The point of building so many nearly empty sites was not to win any of them an audience. It was the aggregate. A claim that appears on a single obscure blog is nothing. The same claim echoed across hundreds of coordinated domains starts to look, to anything scanning the web at scale, like something a lot of independent sources are saying.
The audience was never human
This is where Portal Kombat departs from the older model of Russian information warfare. A troll farm wants engagement. Portal Kombat’s sites often recorded almost none. In one later case documented in Moldova, a network aggregator had published more than 151,000 articles and received a grand total of 86 visits in a month. By the ordinary metrics of media, that is a failure so complete it looks like a mistake.
It was not a mistake, because the readers the network cared about were not people. They were the automated systems that now sit between the public and information: search crawlers, the editors of open reference sites, and above all the large language models behind AI chatbots. A model or a crawler does not weigh whether a source is popular or trusted the way a person does. It ingests text. Flood the web with enough coordinated text repeating the same claims, and you raise the odds that those claims surface when a machine assembles an answer.
Grooming the chatbots
The strategy has a name that researchers borrowed from a darker context: LLM grooming, the deliberate seeding of the open web with content in the hope of shaping what AI language models later reproduce. In February 2025 the American Sunlight Project raised the alarm that the Pravda network’s output, running to millions of articles, appeared engineered less to inform readers than to saturate the training and retrieval data that AI systems draw on.
The follow-up findings were stark. NewsGuard tested leading AI chatbots against false claims traceable to the Pravda ecosystem and reported that the systems repeated the network’s disinformation roughly a third of the time, treating the laundered claims as if they were sourced fact. Investigators at CheckFirst, working with DFRLab, then documented a second vector: the network’s links had been threaded into Wikipedia. They counted more than 1,900 Pravda-network hyperlinks across 44 language versions of the encyclopedia, hundreds of them clustered in articles about the war in Ukraine. Wikipedia matters here for a specific reason. It is one of the most heavily weighted sources that AI models learn from, which means a citation planted in Wikipedia can travel onward into the answers a chatbot gives long after any human editor has moved on.
Why this is harder to fight
Every previous form of this problem had a surface you could push against. A troll account can be suspended. A cloned news site can be seized. A bot network can be taken down, even if it rebuilds. Portal Kombat is diffuse by design. There is no flagship publication to discredit, no single account to ban, no personality to expose. There is only volume, spread across hundreds of interchangeable domains that cost almost nothing to spin up and can be replaced faster than they can be removed. The EU’s disinformation monitors have described this as opening a new front, and the phrase is accurate. The target has shifted from the reader to the retrieval layer, from persuading a person directly to poisoning the well that people increasingly drink from without noticing.
Outcome and what changed
Portal Kombat was named and mapped, and yet it has only grown. Since VIGINUM’s first report the network has expanded into dozens more countries and languages, and the 2025 findings from NewsGuard, the American Sunlight Project, and CheckFirst confirmed that its real target had moved to the AI layer. What changed is not that the network was defeated but that its purpose was finally understood. The defensive tools have not caught up. CheckFirst noted that over three years, community fact-checking on X had addressed the Pravda ecosystem only a tiny number of times, while the network’s own promotional traffic dwarfed those corrections many times over.
That gap is the lasting lesson. Fact-checking is built to answer claims that reach human audiences, and it is poorly suited to a network whose whole purpose is to reach machines, because a false claim that no person reads still cannot be flagged by a person, and yet it can still end up in a model’s output. Portal Kombat gave up on the thing propaganda has always chased, which is an audience, and bet instead that the systems now mediating knowledge would do the distribution for free. So far that bet is paying off, and the operation is still running inside the tools people use to find out what is true.
Leave a Reply