Yandex, Russian Bots and the Evidence Behind the Outage

A line on a traffic chart fell sharply. Within days, it had become a story about the disappearance of Russian propaganda from Western social media.

The appeal was obvious: damaged data centres, silenced bots, a less hostile internet. For a business trying to distinguish organised manipulation from genuine customer anger, it offered something unusually reassuring. Perhaps the noise had a physical switch, and somebody had turned it off.

The evidence available on 15 October 2026 supports a narrower conclusion. Yandex suffered a substantial infrastructure disruption following reported Ukrainian strikes. Much of the disappearing traffic was search crawling. A separately observed stream of malicious automation moved between hosting providers. The claim that Russian social media influence operations had consequently collapsed remains unproven.

That distinction matters beyond the war. A UK retailer deciding whether to relax fraud controls, a communications director explaining a sudden change in sentiment, and an editor commissioning an AI investigation all face the same problem: a measurement can be accurate while the story attached to it is wrong.

How a network outage became a political explanation

BBC Verify's investigation, published on 13 October, reported that three Yandex data centres had been struck since 8 October. It described disruption to services and the circulation of claims that pro-Kremlin bot activity had fallen alongside network traffic.

The consequential step was not recognising an outage. It was identifying what had supposedly gone offline.

In ordinary conversation, a bot often means an inauthentic social media account. In internet measurement, it can mean a programme fetching a page, checking a service or requesting an image. The same word covers activities with very different purposes. A crawler indexing a news article is not equivalent to an account posting a fabricated story about that article.

The ambiguity allowed a chart about requests to websites to become apparent evidence about political participation. Claims about reduced activity around particular movements then supplied a recognisable human narrative for an otherwise technical event.

Snopes, in its report published on 13 October and updated on 14 October, found those causal claims unproven. It described a circulated graph with no source specifying what it measured, alongside other comparisons that did not establish a connection to Russian propaganda accounts.

This is not a reason to assume Russian influence operations are imaginary. Nor is it evidence that right-wing opinion is inherently artificial. Genuine political speech, organised campaigning, commercial spam and covert state-linked manipulation are different categories. Attribution has to be established independently of whether the opinion is popular, objectionable or convenient to an observer's argument.

BBC Verify also reported that President Volodymyr Zelenskyy's office characterised his remarks as describing the atmosphere on social media rather than a measured fall in bot activity. An impression about a feed and a finding about a network are not interchangeable forms of evidence.

For businesses, the reporting lesson is immediate: preserve the distinction between what happened, what somebody said happened, and what a dataset can actually demonstrate.

What Cloudflare measured

Cloudflare's technical analysis of 14 October separates two networks associated with Yandex. AS13238 supports Yandex Search; AS200350 supports Yandex.Cloud. An autonomous system number, or ASN, identifies a network for internet routing. It does not identify a single customer or establish who controls every request originating there.

This separation changes the interpretation of the outage. Search infrastructure and rented cloud infrastructure can both produce automated traffic without serving the same purpose.

A crawler is not a propaganda account

On Yandex Search, Cloudflare classified 99.5% of requests as automated and 93.6% of bot requests as verified. Its inspection found activity consistent with indexing pages, rendering pages and retrieving images.

The denominator matters. The 93.6% figure describes bot requests on the search ASN. It is not the percentage of all Russian automation that is benign, and it says nothing directly about how many Russian-linked social accounts exist.

Cloudflare concluded that the decline on this network was almost entirely search crawling. That is a substantial operational change: publishers may receive fewer indexing requests. It is not, by itself, a reduction in political persuasion, harassment or fabricated engagement.

Nor should a dramatic percentage on a provider-specific chart be mistaken for a global total. A steep fall within one network can coexist with little movement in the broader population. The chart may faithfully show a local collapse while answering none of the questions being asked about international social media.

Global shares need careful handling

Cloudflare's table places global bot requests at 39.4% of all HTTP requests before the event, with readings between 39.2% and 39.6% during it. Those are shares within Cloudflare's observed traffic, not a census of internet accounts.

A percentage is also not an absolute volume. If automated requests and human requests both change, the share may remain similar. Even a stable absolute total could conceal a fall in one activity offset by a rise in another.

The useful conclusion is therefore bounded: the published aggregate does not support a sweeping collapse in automation across the observed network. It cannot independently establish that every influence operation continued unchanged.

There is an additional reason to read the source rather than repeat a detached number. Cloudflare's introductory baseline discussion and its later global-request table display different baseline percentages. The figures used here come specifically from the table labelled as total HTTP requests. They should not be combined with another denominator or presented as an estimate of social media manipulation.

A business dashboard deserves the same scrutiny. Before comparing two percentages, establish whether both count the same event, over the same period, from the same population. Otherwise, precision merely disguises a mismatch.

Malicious traffic moved, but this was not a propaganda finding

Yandex.Cloud presented a different picture. Cloudflare observed reductions in activity resembling vulnerability scanning and probing, including requests looking for WordPress plugin weaknesses. It also described legitimate application traffic. Being unverified as a crawler does not automatically make every cloud-hosted request malicious.

The clearest migration example concerned a gambling customer, not political posting. Cloudflare reported that bot requests from Yandex.Cloud to that account stopped after the disruption, while matching requests appeared from other Russian hosting networks, including Radist and Selectel.

The comparison drew on multiple characteristics: the same JA4 TLS client fingerprint, unusual request paths and headers, and custom User-Agent strings. Taken together with timing and destination, these supported continuity of the observed automation across providers. A fingerprint describes technical characteristics of a connection; it is not a person's identity or proof of state sponsorship.

Across the gambling accounts in its analysis, Cloudflare reported Russia-origin bot requests moving from 167.7 million to 170.1 million, an increase of 1.4%. That observation undermines a claim that this particular automation simply disappeared with the hosting outage.

It does not identify the gambling traffic as a Russian influence operation. It does not demonstrate how a propaganda network would behave. The connection is a resilience lesson, not a shared mission: software can move even when the infrastructure previously running it is unavailable.

For a UK security team, that distinction is more useful than the dramatic headline. A fall in requests from a troublesome hosting provider may represent relief, relocation or an unrelated change in traffic. Check what happens at the targeted service, rather than declaring success because one source network becomes quiet.

The limits of the social media evidence

Cloudflare did not stop at aggregate requests. It reported finding no unexpected changes in its analysis of social media traffic, either globally or from Russian-origin IP addresses. It also examined bot traffic reaching posting, replying and liking endpoints for one particular social platform, again without finding an unexpected shift.

That is relevant evidence against the broad story. It is also a limited observation.

Cloudflare sees requests within its own network and the classifications available to it. The analysis is not an account-by-account audit of every social platform, every language or every covert campaign. An endpoint request is not necessarily a successfully published post; a published post is not necessarily an impression; an impression is not necessarily persuasion.

Cloudflare explicitly said it could not conclusively rule out a connection without further analysis using data outside its network. Recasting its conclusion as proof that no influence activity anywhere was affected would repeat the original mistake in reverse.

The independently reported monitoring evidence is similarly specific. BBC Verify said dTeam shared evidence that Storm-1516 and Matryoshka remained active and reported no traffic change among the known Russian accounts and bot clusters it tracked. That weighs against the proposition that known operations had been broadly silenced. It does not establish complete visibility over unknown operations.

Snopes likewise left the overarching causal claim unproven rather than claiming to have measured every possible effect.

The strongest defensible judgement is neither dramatic success nor universal immunity: available evidence does not substantiate the claim of a major collapse in Russian social media manipulation following the Yandex disruption. More detailed evidence could revise that judgement. Confidence should follow the quality and coverage of the evidence, not the attractiveness of the conclusion.

What would establish a causal connection?

An outage followed by quieter discussion is a starting observation, not a completed investigation. To attribute that change to a disrupted influence operation, an analyst would need to connect infrastructure, accounts, timing and outcomes.

First, define the outcome. Is the claim about the number of active accounts, the volume of posts, coordinated amplification, authentic audience reach or changes in opinion? Each requires different evidence. Counting web requests cannot silently substitute for any of them.

Then establish which accounts belong to the alleged operation, using evidence independent of the outage. Shared political views are insufficient. Relevant corroboration might include platform investigation findings, demonstrable operational links or other records connecting accounts to the same controller. The standard should be higher before naming an organisation or alleging state direction.

Next, demonstrate a dependency on the affected infrastructure. Hosting a related website is not the same as running account orchestration there. An account can remain accessible while its controller loses a service, and a service can fail without affecting every customer. A credible explanation has to show the mechanism, not merely draw a line between two familiar names.

The comparison period also needs care. Account activity should be assessed against a suitable baseline, allowing for posting schedules, major news, moderation decisions and changes in access to data. Comparable unaffected groups or platforms can help distinguish a specific disruption from a wider fluctuation. There is no universally sufficient number of days; the design must fit the claim.

Finally, look for corroborating sequence. Did independently identified accounts fall silent when the relevant service failed? Did activity return when the dependency recovered, or reappear elsewhere with evidence of continuity? Did platform records support that pattern? Such observations would strengthen attribution without requiring public disclosure of private user data.

For a business commissioning a monitoring report, these questions make a useful acceptance test. Ask the supplier to identify the unit of measurement, explain the baseline, separate observation from attribution and describe what evidence would change its assessment. A report that cannot state its limitations should not determine a public accusation.

Where ChatGPT fits, and where the evidence stops

The outage story also invites a second shortcut: if propaganda is automated, and AI writes text, perhaps ChatGPT is its central engine. That conclusion is not supported by the permitted evidence.

OpenAI's disclosure of 30 May 2024 documents actual misuse of its models by covert influence operations. It describes the Russia-origin Bad Grammar operation using models to debug code for a Telegram bot and produce short political comments in Russian and English.

It also describes people acting for Doppelganger generating multilingual comments, translating and editing articles, writing headlines and converting articles into Facebook posts. These are concrete examples of AI supporting production and technical work.

OpenAI said that, as of May 2024, none of the five operations in its report appeared to have meaningfully increased audience engagement or reach because of its services. The five included operations from different countries; they were not five Russian networks.

This is a finding about those investigated campaigns, that provider's visibility and that period. It is not evidence that AI can never improve an influence operation, or that all later attempts failed. Equally, documented use of OpenAI models does not establish ChatGPT's dominance across Russian propaganda. A supplier's investigation is not a market-share survey of covert activity.

The disclosure also says these operations mixed AI-generated material with other formats rather than relying exclusively on AI. That is the more useful way to think about the problem. Content creation, account control, publication and audience distribution are separate activities, even when one operator combines them.

A human can circulate AI-written material manually. Software can distribute human-written material automatically. A deceptive campaign can use authentic articles selectively without generating new prose at all. None of these arrangements can be inferred simply from the label attached to a traffic chart.

Prose cannot identify the model or the sponsor

A fluent paragraph is not evidence that ChatGPT wrote it. Awkward English does not prove a foreign operator, and repeated wording does not by itself prove automation. Templates, translations, quotations and coordinated but legitimate campaigning can all produce similarities worth investigating without resolving attribution.

For this article, no model identity is inferred from writing style. A reliable claim about the use of a named model would require corroboration such as provider findings or authenticated operational records, not a reader's intuition about the prose.

The same discipline applies to internal AI investigations. Do not ask a chatbot to turn suspicious wording into a confident nationality, sponsor or product attribution. Ask it to organise the evidence and identify unanswered questions. Keep any proposed attribution explicitly provisional until a person has checked the underlying material.

A security response for UK businesses

The immediate defensive implication is not to lower the alert level because a foreign provider has an outage. It is to check whether harmful activity against the business has changed.

A small security team can begin with the services it actually operates: customer login, account recovery, payment journeys, public forms and any administrative interfaces. Review unusual request patterns alongside failed authentication, affected accounts, service errors and customer reports. A lower traffic total is reassuring only if the harmful outcome also falls.

For an outsourced service, ask the provider which observations it can supply. An ASN breakdown may help locate traffic, but a useful incident update should also describe the targeted function, observed behaviour, mitigation and impact on legitimate customers. Avoid requesting unnecessary personal data merely to make the dashboard look complete.

Cloudflare's migration finding supports checking continuity across changing source networks. It does not justify indiscriminate blocking of all traffic from a country or treating a shared fingerprint as conclusive proof. Controls should be proportionate, reviewed for false positives and tested against genuine customer journeys.

Our discussion of UK SME cyber defence in the AI arms race provides a related security perspective. The Yandex lesson is more specific: keep the observation of malicious requests separate from assumptions about the actor's political purpose.

There is also a useful architectural question. Which public-facing functions genuinely need a dynamic application, and which could remain available as simpler published pages? The analysis of WordPress exposure and static deployment explores that choice. It is not a claim that static publishing prevents disinformation or removes every security risk; it is a prompt to examine which dependencies a service actually requires.

Record what would justify relaxing any emergency control before relaxing it. That might include a sustained reduction in confirmed abuse and no increase through another route. Provider silence alone is not the outcome the business is trying to achieve.

Communications without accidental amplification

A communications team faces a different hazard: explaining activity that the security team has not attributed. Public criticism can be authentic, coordinated, automated or a mixture. Treating all criticism as hostile interference risks dismissing customers while giving an unsupported allegation a much larger audience.

For an unexplained surge, preserve relevant public posts, their URLs and collection times. Distinguish original material from reposts and distinguish engagement counts from unique participants. Keep a note of how monitoring data was collected, because a changed query or interrupted feed can create an apparent change in sentiment.

Escalation should follow business impact and the content of the threat. Impersonation, a false safety claim and ordinary political disagreement require different responses. Coordinate with security where there are suspicious links or compromised accounts, but do not convert a technical indicator into a public geopolitical accusation.

A holding statement can describe verified facts without identifying an unproven sponsor: what customers should rely on, which official channels remain authoritative and what the organisation is investigating. If false material is circulating, correct the relevant factual claim and give readers a direct route to check it. Repeating an inflammatory rumour in every response can make it the centre of the story.

Internal briefings need the same restraint. Label direct observations, reported findings and working hypotheses separately. A senior audience does not need every log entry, but it does need to know whether a conclusion rests on platform evidence, third-party monitoring or a screenshot circulating online.

Agree who can approve attribution and corrections before an incident. The pressure to sound decisive is strongest precisely when evidence is incomplete. A clear approval route lets the business respond without pretending to know more than it does.

Media literacy starts with the commissioning desk

The Yandex episode is also a warning about how a plausible research document can become a confident article. Named organisations, dense technical language and precise percentages do not make every connecting sentence reliable.

Editors should open the cited source and ask whether it supports the specific claim being published. A technical report can establish a reduction in scanning without supporting the adjacent claim about political influence. A fact-check can describe an allegation without endorsing it. A source quoted by several outlets does not become several independent datasets.

The practical editorial safeguard is a short claim record: the proposed assertion, its direct source, the relevant date, the population measured and the caveat that must remain attached. This need not become a bureaucratic exercise. It is a way to prevent the most important limitation from disappearing during rewriting.

For an AI-assisted draft, check links before assessing polish. A model can join supported facts with an unsupported causal bridge that reads perfectly naturally. Human review should therefore inspect the joins, not just spelling and grammar. The question is not merely whether each number appears somewhere, but whether those numbers justify the conclusion.

The discussion of AI content and human editorial oversight is relevant here as further reading, not evidence for the outage. Regardless of how a draft was produced, the publisher owns the decision to present a claim as fact.

Apply the standard consistently. If a story flatters the reader's political preferences, it still needs attribution and a defensible method. Media literacy is least useful when reserved for information one already dislikes.

Local infrastructure does not settle information trust

It is tempting to turn the story into a procurement argument: domestic hosting is fragile, overseas hosting is resilient, or a local model is inherently safer than a cloud service. The cited evidence cannot sustain those general conclusions.

Hosting location, service dependency and content trust are separate design questions. A local installation may change who operates the infrastructure; it does not make every imported document accurate. A cloud service may offer useful controls; it does not make every generated answer reliable. An organisation needs to evaluate its own requirements rather than inherit a geopolitical explanation from a news story.

For that discussion, sovereign AI and local language models and the distinction between air-gapped AI and instructions about data access address related choices. Neither should be read as evidence that a particular deployment arrangement prevents propaganda.

A sensible procurement review asks what the service depends on, what happens during disruption, how access is controlled and how unsupported outputs will be caught. Test the contingency that matters to the business. A model being locally available and an answer being editorially defensible are different forms of assurance.

Sources and method

This article assesses public evidence available on 15 October 2026. It does not independently verify physical strike damage, identify covert accounts, analyse private platform records or infer a model from its output. Business recommendations are editorial analysis, not reported findings about a specific UK organisation. The linked internal articles are contextual reading, not corroboration for the factual claims about Yandex.

Cloudflare, 14 October 2026: Yandex outage analysis supplies the ASN distinction, verified-crawler figure, global-request table, gambling-traffic migration and limited social-platform endpoint findings. Its coverage is traffic observed across its network; its explicit uncertainty about a wider connection is retained here.

BBC Verify, 13 October 2026: investigation of the bot-disruption claim supplies reporting on three affected sites, dTeam's observations of known networks, Yandex's stated policy against manipulation and the response from Zelenskyy's office. A stated policy is not evidence that every customer complied with it.

Snopes, published 13 October and updated 14 October 2026: assessment of the causal claim supplies the finding that the claim remained unproven and the examination of circulated charts and comparisons. Its conclusion is not treated as a census of all influence activity.

OpenAI, 30 May 2024: disruption of covert influence operations documents misuse by Bad Grammar and Doppelganger and reports no meaningful increase in reach or engagement attributable to its services across the five investigated campaigns at that time. It is a provider disclosure, not evidence of ChatGPT's prevalence across all Russian operations.

Conclusion

The Yandex disruption was real enough to generate visible network effects. The story that those effects represented a collapse of Russian propaganda went further than the evidence allowed.

For UK businesses, the answer is not indifference to influence operations or complacency about AI misuse. It is better separation: requests from accounts, accounts from audiences, production from persuasion, and observation from attribution. Protect services against the behaviour reaching them. Explain verified facts without laundering speculation into certainty. Demand the original evidence before allowing a persuasive chart to become a business decision.