Ukrainian Drone Strikes on Yandex Data Centers: Effect on Online Bots

•

 36 min video

•

 13 min read

YouTube video ID: JnceIDVMZyQ

Source: YouTube video by Patrick Boyle — Watch original video

PDF

On October 11th, a chart posted on Twitter by a Member of the European Parliament claimed that online activity related to Germany's AFD party had plummeted after Ukrainian drones struck Russian data centers. This chart, which garnered over 4 million views, was notably convincing despite including data from three days in the future. While the poster did add a disclaimer "if confirmed," the chart itself lacked crucial details such as a data source, the specific online platform it covered, and what "activity" it was measuring. Its title, "daily activity over 3 months," was vague, and it didn't even mention the AFD. Furthermore, the scaling on both axes was manipulated to make a decline appear steeper. German magazine Capital traced the chart to a Threads account purportedly belonging to a 24-year-old flight attendant, and German broadcaster DW noted that while AI involvement couldn't be proven, it couldn't be ruled out either.

The appeal of such a narrative was understandable, especially given the British Prime Minister's recent comments about Russian interference in the Brexit referendum and the upcoming US midterms. The idea that drone strikes could effectively neutralize foreign propaganda and improve social media seemed like a decisive solution to content moderation issues.

The Reality of the Drone Strikes

In the preceding month, Russia had been targeting data centers in Ukraine, a move the Russian Defense Ministry termed a "digital detox." Ukraine retaliated by striking three of the six major data centers operated by Yandex, often referred to as Russia's Google. Yandex also manages Russia's largest taxi app, a food delivery service, a payments business, and an AI chatbot. Two of the targeted sites were completely shut down, and a third lost several modules within hours.

Following these strikes, charts from Cloudflare Radar showed a significant drop in Russian bot traffic. Euro News estimated that automated traffic from Yandex's network had decreased by approximately 94%.

Cloudflare, an American company that acts as an intermediary between websites and their visitors, protects sites from attacks and distinguishes between human and automated traffic. Its public dashboard, Radar, provides insights into internet traffic. Cloudflare defines a bot as any automated online request. Before the strikes, bots accounted for over a third of all traffic from Russia handled by Cloudflare. Most of this traffic consisted of search engine crawlers and other scripts, rather than malicious online trolls. Cloudflare confirmed that the traffic that collapsed was almost entirely Yandex's search crawlers, such as Yandexbot, which index news websites for the search engine.

Essentially, the internet had celebrated the destruction of a search engine's "filing clerk," a bot that about 40% of the news sites it visited had already blocked. Yandex's entire search network constituted only about 0.05% of the world's bot traffic. Cloudflare then examined traffic to social media sites, including posting, replying, and liking activities from Russia, and found no unexpected changes. While Cloudflare could only report on traffic within its own network, it concluded that troll farms had not missed any work.

Collateral Damage and Russia's Data Localization Policy

Despite the limited impact on social media bots, the drone strikes caused significant damage. According to the Financial Times, users of Yandex's payment app found their balances had disappeared. Bus timetables vanished in some cities, and residents of many Moscow apartment blocks were locked out of their courtyards because barrier gates relied on Yandex services.

This extensive damage from just three strikes can be traced back to a decision made by Vladimir Putin in 2014. Putin, who once described the internet as a CIA project, signed a law that year requiring Russian citizens' personal data to be stored on servers within Russia. This data localization law led to LinkedIn being blocked in 2016 for non-compliance.

Initially, this policy seemed advantageous. When Western cloud providers like Amazon, Microsoft, and Google ceased operations in Russia after the invasion of Ukraine, the Russian internet continued to function because it was run by Russian companies on Russian servers, insulated from foreign sanctions. However, the downside was that a significant portion of Russia's digital economy became concentrated in a few buildings within drone range of Ukraine, making critical infrastructure vulnerable. The data localization law also prevents Russian companies from moving their data to safer locations abroad. Consequently, Yandex advised customers to back up their data with rival providers, most of whom also have data centers near Moscow and St. Petersburg.

Financial Impact and Recovery Challenges

Investors reacted relatively calmly to the strikes, with Yandex's shares falling by about 11%. This modest reaction is partly because the cloud business, the directly affected segment, accounts for only about 4% of Yandex's revenue. The company's primary revenue streams come from taxis, advertising, and online shopping, which could potentially be shifted to remaining data centers.

However, Yandex's customers faced greater challenges. Thousands of Russian businesses renting space on Yandex's servers were severely impacted. One Russian telecom analyst estimated daily losses in the billions of rubles. Recovery will be slow; Yandex hopes to restore services quickly, albeit at a slower pace, but reaching full capacity could take several months. Replacing servers will be difficult due to Western sanctions restricting computer equipment to Russia, likely necessitating purchases from China or through intermediaries in Central Asia, which is slower and more expensive.

The responsibility for protecting this critical infrastructure is a point of contention among Russian officials. The State Duma places the onus on companies, while Putin has reportedly told businesses not to expect state support. The Digital Ministry, however, claims to be working with the Defense Ministry on protection. This ambiguity is likely not reassuring for data center operators.

Interestingly, one firm selling mobile data centers to Russian companies reportedly has a deputy director whose father is the first deputy director of the FSB, suggesting potential profiteering from the situation.

Redundancy and Data Sovereignty

Standard practice for cloud companies to protect against data center outages is redundancy, distributing work across multiple sites. Yandex's own documentation outlines proper redundancy, recommending that if an application needs nine virtual machines, nine should be in one location and nine in another, totaling 18 machines. This acts as an insurance premium, with customers paying for capacity they hope not to use. However, customers who followed this manual and split their servers between two locations might have found that Ukraine hit both. After the third strike, Yandex stopped selling new computing capacity, leaving customers who hadn't paid the premium without options. With half of its major data centers out of action, remaining customers have no alternative if another strike occurs.

The Yandex situation highlights the complexities of data sovereignty debates in Europe and Canada. While often framed as a national security issue, building domestic cloud infrastructure to avoid reliance on foreign tech giants also presents a risk management problem: balancing dependence on a single foreign supplier against the physical vulnerability of keeping all servers domestically. Even American cloud providers are not immune; Iranian strikes on Amazon's data centers in the Middle East resulted in permanent data loss for customers in the UAE and Bahrain. The lesson from Yandex is to diversify data storage and ensure each location has sufficient capacity.

The True Source of Online Influence

If trolls weren't primarily in Yandex's data centers, where were they? Russian influence operations often involve buildings full of people, like the Internet Research Agency in St. Petersburg. However, to avoid easy detection and blocking, these operations rent internet connections.

Online influence campaigns have two main components: the accounts that post content and the websites they link to. Websites are easier to manage. Russia's "Doppelganger" campaign, for instance, has been cloning Western news sites since 2022, creating fake versions of Der Spiegel, Le Monde, and The Washington Post on lookalike web addresses filled with fabricated stories. An investigation by Correctiv traced part of this operation to a data center run by German hosting company Hetzner near Helsinki. Hosting companies generally do not fact-check their customers.

User accounts are more challenging. Social media platforms identify traffic from data centers as likely automated. Serious bot operations bypass this by using residential connections and phone farms, which are not susceptible to air strikes. Residential connections are home internet lines, often rented out without the homeowner's knowledge. In May 2024, the US Justice Department dismantled 911S5, a botnet that infected millions of home computers through free programs and pirated software, then rented out their internet connections for fraudulent activities. Similarly, Google disrupted "EPIDA," a network that turned devices into relays through Android apps and Windows programs, used by various threat groups for activities like creating fake accounts.

Phone farms address the need for multiple phone numbers to create thousands of social media accounts. Europol raided a SIM farm in Latvia, seizing 1,200 devices with 40,000 active SIM cards, which had been used to open over 49 million online accounts.

The Role of Off-Brand Android TV Boxes

A significant, often overlooked, component of the modern disinformation economy is the inexpensive off-brand Android TV box. These devices, bought online and connected to TVs and Wi-Fi, run insecure, stripped-down Android versions. Many come with pre-installed malware, allowing them to be easily compromised. The "Kim Wolf" botnet infected over 2 million such devices, with EPIDA's infrastructure supporting it. Millions of these boxes stream content while secretly renting out their owners' internet connections. The takeover process is often simple, exploiting default manufacturer usernames and passwords like "admin."

Government-Sponsored Bot Farms and "Rage Bait"

Once these hijacked devices and SIM cards are assembled, customers are needed, and governments are often reliable clients. In July 2024, the US Justice Department shut down a Russian bot farm of nearly a thousand Twitter accounts created with AI software called "Melorator." This project, pitched by a deputy editor-in-chief at RT, was later integrated into an FSB-backed organization. The software generated fake personas and mimicked ordinary social media users. When the FBI seized the domains used for email addresses, the bots lost their ability to bypass Twitter's filters.

State interference is a documented reality, though its impact on actual votes is harder to quantify. Meta and OpenAI have reported that Russian AI-driven influence campaigns largely fail to reach large audiences, yet governments continue to fund them. This phenomenon is not exclusive to Russia; a 2019 Oxford Internet Institute report found organized social media manipulation in 70 countries, with campaigns traced to governments including China, Iran, and Saudi Arabia. In 25 countries, governments hired private firms for "computational propaganda as a service." Western democracies also engage in such activities; the Oxford Report cited a GCHQ unit that created Facebook groups and YouTube videos to "discredit, promote distrust, dissuade, deter, delay, and disrupt." A US aid program even built a fake social network in Cuba, "Cuban Twitter," which aimed to introduce political content after building an audience.

The Economics of Political "Rage Bait"

Beyond government actors, many influential political voices on the internet are driven by financial incentives rather than ideology. Last November, Twitter's feature showing an account's country of origin revealed that many highly patriotic American political accounts were operated from countries like Nigeria, Bangladesh, and Eastern Europe. This wasn't politically one-sided; a "Republicans against Trump" account was found to be operating from Austria.

The motivation behind this is advertising rates. A report by the Institute for Strategic Dialogue, "Hate for Profit," identified 128 Facebook pages and groups posting anti-migrant content aimed at British audiences, all traced to one man in Sri Lanka. He runs an online academy teaching students to profit from political "rage bait" on Facebook, explaining that politics is the niche most likely to go viral. The economics are simple: a thousand views from a Pakistani audience might earn a creator 40 cents, while the same views from the US, UK, or Canada could earn $6 to $30—15 to 75 times more for the same effort. Western political anger has become a high-margin export from emerging markets, where accuracy is secondary to engagement.

Platforms contribute to this by sharing ad revenue on videos, photos, and text posts. The more engagement a post gets, the more it earns. Twitter's attempt to prioritize views from a creator's home region to curb this was quickly paused after user complaints.

The Human Element in Disinformation Spread

While bots play a role, research indicates that real people do most of the heavy lifting in spreading disinformation. Mark Zuckerberg's "natural engagement pattern" suggests that unchecked, people disproportionately engage with sensational and provocative content. Facebook's decision to pay creators based on engagement, despite its stated aim to reduce borderline content, incentivizes this behavior.

A 2018 MIT study of 126,000 Twitter stories between 2006 and 2017 found that false news spread further and faster than truth, especially in politics. False stories were 70% more likely to be retweeted and reached 1,500 people six times faster. Even after removing bots, this gap persisted. The researchers attributed this to novelty; false stories elicited more surprise and disgust, which are more engaging than the sadness, anticipation, and trust associated with true stories.

Furthermore, a 2021 study in Science found that a small percentage of "super sharers"—just 0.3% of 660,000 registered American voters on Twitter—were responsible for 80% of the fake news shared during the 2020 election. These were ordinary people, averaging 58 years old, manually retweeting almost 16 political news links daily.

Often, the most damaging content isn't fake. During the early COVID vaccine rollout, fact-checked misinformation accounted for only 0.3% of 2.7 billion views of vaccine-related links. While it reduced vaccination willingness, few people saw these posts. The most viewed story, a Chicago Tribune article about a healthy doctor dying after a COVID vaccine, was seen by over 50 million people on Facebook and garnered six times more views than all flagged misinformation combined. Overall, unflagged vaccine-skeptical stories had 46 times the effect on vaccine hesitancy.

This brings us back to the initial chart: it told people what they wanted to hear, leading to widespread sharing despite its flaws.

The Future of Self-Replicating AI and Defense Mechanisms

The ease with which devices are compromised is striking. The Mirai botnet, for example, spread by exploiting default manufacturer usernames and passwords, often "admin." Researchers at Anthropic recently tested whether AI models could run on and spread across off-brand Android TV boxes. They found that an AI model running on 15 such devices could process 0.39 tokens per second, roughly one tweet every three minutes, making it a productive account. The paper warns that small AI models could soon self-replicate across such devices. In simulated tests, the hacking step was trivial due to a lack of passwords. Given how Mirai spread, real-world security might not be much harder to overcome. The researchers likened this to a "slime mold"—slow but difficult to shut down due to its decentralized nature, the opposite of Yandex's problem.

The defense against such self-replicating digital slime molds is surprisingly simple. One internet provider quarantined infected customers and informed them, clearing 92% of Mirai infections within 14 days. The UK has also made default passwords like "admin" non-compliant for devices sold in its markets. Humanity's first line of defense against self-replicating AI may be to urge manufacturers to stop using "admin" as a default password.

The researchers also found that small AI models are "very poor to poor" at the persuasion aspect of hacking, meaning they struggle to trick people with convincing messages. So, even if a self-replicating AI takes over a television, it would still need ordinary people to furiously retweet its posts for them to gain traction.

The off-brand Android TV box has evolved from a device for streaming pirated football to a residential proxy, a patriotic American voter, and potentially a host for a digital slime mold—quite a journey for a $55 device.

  Takeaways

  • The viral chart claiming a sharp drop in AFD‑related online activity after Ukrainian drone strikes was misleading, using future data, vague metrics, and manipulated axis scales, and lacked source information.
  • The drone attacks destroyed three Yandex data centers, cutting off Yandex’s search‑crawler traffic, which accounted for about 94% of the observed decline in Russian bot traffic on Cloudflare Radar, but had minimal impact on social‑media activity.
  • Russia’s 2014 data‑localisation law concentrated critical digital services in a few Moscow‑area facilities, making them vulnerable to physical attacks and limiting companies’ ability to relocate data abroad.
  • Yandex’s cloud segment represents only about 4% of its revenue, so its share price fell modestly, yet thousands of Russian businesses suffered billions of rubles in daily losses and face a slow recovery due to sanctions on replacement hardware.
  • Disinformation campaigns rely more on human “super‑sharers” and rented residential or SIM‑farm connections than on data‑center bots, and the proliferation of insecure off‑brand Android TV boxes creates a new vector for self‑replicating AI malware.

Frequently Asked Questions

Why did Cloudflare Radar show a 94% drop in Russian bot traffic after the drone strikes?

Cloudflare Radar recorded the decline because the strikes disabled Yandex’s search‑crawler servers, which generated the majority of Russian automated requests; with those bots offline, the measured bot traffic fell sharply. The drop reflected Yandexbot loss, not a reduction in human or troll activity on social platforms.

What is Russia’s 2014 data‑localisation law and how did it make Yandex’s data centers more vulnerable?

The 2014 law requires personal data of Russian citizens to be stored on servers located within Russia, forcing companies like Yandex to keep critical services in a few domestic facilities; this concentration meant that a few Ukrainian drone strikes could cripple major online services and left no easy option to relocate data abroad.

Who is Patrick Boyle on YouTube?

Patrick Boyle is a YouTube channel that publishes videos on a range of topics. Browse more summaries from this channel below.

Does this page include the full transcript of the video?

Yes, the full transcript for this video is available on this page. Click 'Show transcript' in the sidebar to read it.

unit that created Facebook groups and YouTube videos to "discredit, promote distrust, dissuade, deter, delay, and disrupt."

US aid program even built a fake social network in Cuba, "Cuban Twitter," which aimed to introduce political content after building an audience.

Helpful resources related to this video

If you want to practice or explore the concepts discussed in the video, these commonly used tools may help.

Links may be affiliate links. We only include resources that are genuinely relevant to the topic.

Full transcript is not shown on this page

This page focuses on the summary and original notes. For full verification, refer to the original YouTube video.

PDF