Did Ukraine’s Strikes on Yandex Hit Russian Propaganda? New Study

Ukrainian strikes on Yandex data centers coincided with a measurable drop in Russian disinformation campaigns, according to new research. An analysis of 75,000 social media posts reveals a sharp decline in key propaganda narratives following the October attacks, highlighting potential vulnerabilities in Moscow’s digital infrastructure.
The affected topics concerned Ukraine's military mobilisation system, warnings of a utilities collapse in Kyiv and calls for a peace agreement with Russia.
Other narratives continued circulating. Some had already lost momentum before the strikes.
The findings raise an important question for governments confronting Russian information operations: can attacks on digital infrastructure disrupt organised propaganda campaigns, or do the apparent changes simply reflect the unpredictable behaviour of online audiences?
CAT-UA's researchers believe a connection is plausible. Their evidence, however, is indirect.
From Viral Claims to Measurable Evidence
The controversy began on 11 October, the day of the second reported strike against Yandex data centres.
A graphic circulated online purporting to show Russian bot activity collapsing to zero. Ukrainian fact-checkers challenged its authenticity, pointing out that no publicly available monitoring system could reliably measure the activity of every Russian bot network in real time.
The graphic was misleading, but dismissing it did not answer whether the strikes had affected actual influence operations.
Ukraine's Centre for Countering Disinformation subsequently published findings based on accounts flagged by the Osavul monitoring system for potentially inauthentic behaviour. CAT-UA decided to conduct a separate investigation, citing limitations in the transparency of the available methodology.
Its researchers examined social media activity between 13 September and 12 October 2026 across Facebook, X, Telegram, YouTube, Threads, TikTok, Instagram and Bluesky.
They selected five subjects already prominent in Russian propaganda: criticism of Ukraine's military recruitment authorities, demands for President Volodymyr Zelenskyy's resignation, claims about collapsing public services in Kyiv, calls for a peace agreement and criticism of Ukrainian weapons manufacturers.
Each topic produced between 64,000 and 140,000 collected messages. Researchers randomly selected 15,000 from each dataset, creating a combined analytical sample of 75,000 posts and comments.
A trained artificial intelligence model classified the messages, while the researchers examined unusually active accounts, frequently shared links and recurring narratives.
The investigation looked for changes around the dates of the strikes rather than attempting to count every automated account.
That methodological distinction is essential. A decline in propaganda-related messages does not necessarily mean an equivalent decline in bot activity.
Why AI Makes Russian Bot Networks Harder to Detect
Identifying coordinated manipulation has become more difficult as artificial intelligence changes the way online campaigns operate.
Earlier methods relied heavily on finding identical or nearly identical messages posted repeatedly across different accounts. Researchers could identify clusters of duplicated text, examine suspicious posting patterns and separate coordinated campaigns from ordinary viral content.
Those techniques remain useful, particularly for less sophisticated operations.
But AI-generated text allows operators to circulate similar arguments using different wording. Networks can also rely on purchased or compromised accounts that resemble ordinary users rather than profiles posting hundreds of messages within an hour.
A campaign may therefore retain a consistent political message without producing the obvious textual fingerprints associated with traditional bots.
CAT-UA used AI-assisted analysis to identify recurring arguments and suspicious patterns across the selected datasets.
Even so, the organisation acknowledges that its approach cannot establish the precise number of automated accounts involved.
There is another complication. People can independently repeat narratives promoted by Russian propaganda without being part of an organised operation.
Criticism of mobilisation, for example, can emerge from genuine Ukrainian experiences and political disagreements. Its appearance online cannot automatically be classified as foreign manipulation.
The researchers therefore treated changes in message circulation as potential indicators of coordinated activity, not definitive proof of automation.
Three Narratives Showed Declining Activity
The clearest patterns appeared in discussions of mobilisation, Kyiv's public utilities and negotiations with Russia.
Messages criticising Ukraine's Territorial Recruitment and Social Support Centres, commonly known by their Ukrainian abbreviation TCC, fell towards monthly lows around 10–11 October.
These centres administer military registration and mobilisation procedures, making them a frequent subject of domestic controversy and Russian information campaigns.
Yet the decline was uneven.
CAT-UA found that some coordinated activity continued, including promotion of the website Busification.org, which actually experienced an increase in coordinated sharing on 11 October.
Other narratives, including claims that mobilisation was coercive, declined more sharply.
The mixed pattern is consistent with disruption affecting certain campaigns while leaving others operational. It could also reflect the normal fluctuations of separate online discussions.
The findings concerning Kyiv's utilities were similarly complicated.
Russian attacks on Ukrainian infrastructure had generated genuine public concern about electricity, water and heating. Social media attention increased during October, driven partly by actual developments and statements from local authorities.
Researchers nevertheless identified declining activity across several recurring narratives on 11 October.
One exception involved claims that Kyiv was facing a sewage catastrophe. That narrative reached a peak on the same day, following a statement by Kyivvodokanal concerning wastewater reception.
The episode illustrates how propaganda can amplify genuine public-service concerns while exaggerating their consequences.
Calls for negotiations with Russia showed a different pattern.
Overall activity and messages arguing that talks with Moscow were necessary began declining sharply on 8 October. But discussion remained above September levels, and coordinated distribution of statements by Russian officials continued.
CAT-UA considers it possible that several networks were promoting the same subject and that only some experienced disruption.
The research does not establish whether those networks depended on Yandex infrastructure.
The Evidence That Complicates the Theory
Two other topics offered little support for a direct connection between the strikes and changing online activity.
Messages demanding Zelenskyy's resignation had already declined sharply around 1 October, before the reported attacks on the data centres.
CAT-UA associated this earlier decline with a shift in political attention following criticism of Zelenskyy by US President Donald Trump.
The researchers suggested that operators may have reduced their activity because the political environment had become less favourable to attacks on the Ukrainian president.
That remains an interpretation rather than a documented decision by any identified network operator.
The fifth subject, criticism of Ukrainian weapons manufacturers, produced fewer signs of potentially Russian coordination. Much of the activity appeared connected to domestic political disputes.
Although overall attention declined after 8 October, the researchers could not establish whether this represented disruption to Russian influence operations.
These exceptions matter because they prevent the study from supporting a simple narrative in which an infrastructure strike disabled a single, centrally controlled propaganda system.
The evidence instead points to different campaigns behaving differently, potentially reflecting separate operators, infrastructure and political objectives.
What the Findings Mean for Europe’s Information Security
For European governments concerned about Russian hybrid operations, the study offers a useful example of both the potential vulnerability and the resilience of coordinated online influence campaigns.
If some networks were disrupted by the loss of data-centre capacity, it would indicate that at least certain information operations depend on technical infrastructure vulnerable to interruption.
But the research contains no direct technical evidence connecting particular accounts to Yandex servers. It does not identify the systems used to coordinate the campaigns or establish how long any disruption lasted.
Consequently, the findings cannot demonstrate that striking a major technology provider is an effective general method of countering disinformation.
They do, however, expose a challenge facing European researchers and institutions attempting to measure hostile information activity.
The increasing use of AI-generated messages makes it harder to distinguish centrally organised manipulation from ordinary political discussion. That problem becomes particularly acute when influence campaigns exploit genuine public grievances.
A credible assessment requires more than counting hostile comments or identifying a sudden decline in traffic. Researchers must examine how messages spread, whether accounts behave in coordinated ways, and what external events might explain changes in public attention.
CAT-UA has made its methodology available for scrutiny and invited other researchers to examine its interpretation.
Its central finding remains deliberately limited: three of five monitored narratives declined around the relevant dates, making disruption of some bot networks a plausible explanation.
The unanswered question is whether future investigations can connect those changes in online behaviour to specific technical failures. Without that link, the October data provide evidence of a possible disruption, but not proof of what caused it.
