Baseline Measurement: How Much Propaganda Is on the Web Without Yandex?

Drone Strikes Lead to Collapse in Russian Bot Traffic

In the early morning of Sunday, October 11, Ukrainian drones hit a data center belonging to the Russian technology group Yandex in the Energetik district of the city of Vladimir. Yandex confirmed on Telegram that the facility’s operations were “completely suspended” and that no one was injured. The facility is designed for 40 to 50 megawatts and up to 2,880 server racks. It was the third strike on Yandex server infrastructure within four days. Ukrainian drone manufacturer Fire Point stated that the attack was carried out with FP-1 drones.

The outage affected almost the entire Yandex portfolio. Consumer services such as the voice assistant Alice, Yandex Music, the Telemost video conferencing tool, and Yandex Smart Home were affected. On the corporate side, Compute Cloud, Object Storage, Managed Kubernetes, as well as PostgreSQL and ClickHouse databases went down, along with developer tools and AI interfaces like the YandexGPT API, SpeechKit, and Vision OCR. The Yandex Cloud status page temporarily listed at least 84 services as offline. Users also reported disruptions in banking apps, ride-hailing platforms, and the retail chains Magnit and Fix Price, some also in neighboring countries.

Cloudflare registered a collapse in bot traffic from the Yandex network to 0.4 percent of the maximum measured in the past two weeks; after the first strike on October 8, the figure was still around 57 percent. The monitoring company NetBlocks puts the remaining network traffic at 30 percent for Yandex Cloud and 63 percent for Yandex LLC. Switching to satellite internet is hardly an option for Russian users, as Starlink is not available there and the competing Rassvet constellation has lost some of its satellites. The governor of the Vladimir region, Alexander Avdeev, reported damaged substations and stated that heating, water, and about 80 percent of the power supply had been restored by Sunday morning. The Russian Ministry of Defense stated that it had intercepted 479 Ukrainian fixed-wing drones over 16 regions during the night.

A viral post on X claims that support for the AfD on social media plummeted by 56 percent following Ukrainian attacks on Yandex data centers. The author cites other alleged observations: fewer supposedly Polish accounts inciting hatred against Ukrainians, fewer supposedly Ukrainian accounts inciting hatred against Poles, and a noticeable decline in online support for MAGA content. The post provides no evidence for these figures. There is no information on the measurement period, the data basis, or whether the figures refer to reach, interactions, or account numbers. Independent confirmation through platform data or research institutes is not yet available, nor has a causal link between the data center outage and the described decline been proven.

This was preceded by two strikes in the same week: On October 8, a Yandex data center in Sasovo in the Ryazan region was hit, causing a fire and forcing a complete shutdown of the facility. It housed server infrastructure for, among others, the Russian railway company RZhD, the mobile provider MegaFon, and several national media outlets. On October 9, a site in Kaluga followed, where several hardware modules failed, according to Yandex. According to Defense Express, Sasovo housed two of Yandex’s three supercomputers, and Vladimir housed the third, “Galushkin,” which are central to training the company’s proprietary AI models; this means all three are now out of operation. After the first attack, Yandex removed the location details of its data centers from its public materials.

Kyiv justifies the attacks as a way to hit Russia’s sources of revenue for its war effort and to force Moscow to divert resources for repairs. Russia had previously shelled Ukrainian data centers, leading to internet outages and prompting some companies to move servers abroad. According to assessments from OSINT channels, the companies most affected are those that do not operate their own IT infrastructure and run their websites, databases, and payment systems entirely on Yandex Cloud; a complete list is not yet available. Sunday’s attack occurred a few hours after a statement by US President Donald Trump that Ukraine should refrain from attacking Russian oil refineries. → Kyiv Post, Tom’s Hardware, BBC, Defense Express, X

Synthszr Take: A single building with space for 2,880 racks was enough to simultaneously paralyze the point-of-sale systems at Magnit and Fix Price, banking apps, and the AI interfaces of an entire country. Digital sovereignty is determined by physical addresses, and three strikes in four days have reduced Russia’s largest cloud to 30 percent of its remaining traffic. For European companies, this poses an inventory question: How many buildings would someone need to know about to halt your own operations? Most can’t name that number, and providers are reluctant to disclose it, which is why Yandex took the locations of its data centers offline after the first strike. Geographic redundancy only costs money in normal operation and is therefore not part of any business case—until the morning it’s the only thing that matters.

Musk to also run Grok Bot on Claude Opus 5.5

On October 7, 2026, Elon Musk announced that the Grok Bot agent will no longer work exclusively on in-house models. The best backend model for each task will be used, with Claude Opus 5.5, MidJourney, and Suno mentioned among others. Grok Bot is the personal agent from SpaceXAI, the AI division integrated into SpaceX; it launched in August, runs on a dedicated machine for each user, and competes with Meta’s Muse and OpenAI’s Dots. Until this announcement, the agent and the model behind it were a single in-house product. A day later, according to The Business Engineer newsletter, Google introduced a single Gemini agent for work tasks that executes each job on the most suitable model. → The Business Engineer

Synthszr Take: Three years and billions spent on in-house models, and now his own agent is calling on Claude Opus 5.5: Musk is publicly admitting that the division of labor between agent and model matters more than the origin of the weights. The hard problem is drawing the line: how much structure an agent needs before it constricts a capable model, and how little is still enough for reliable results. This line shifts with each model generation, which is why SpaceX’s acquisition of Cursor looks smarter than any further training budget.

Trump orders agencies to say “Super Intelligence” instead of AI, Big Tech voluntarily complies

The US government has issued an order directing all federal agencies to write “SI” instead of “AI” in official communications, on websites, and in policy documents, reports Techstrong.ai. On Truth Social, President Trump declared that anyone who continues to use the old term will be considered “THE ENEMY!”. The order is not legally binding for private companies, yet several corporations are following suit: Salesforce CEO Marc Benioff renamed the AIForce agent platform to SIForce, Elon Musk announced the renaming of SpaceXAI to SpaceXSI, and Satya Nadella adopted the new terminology in posts on AI governance. Nvidia CEO Jensen Huang attended the accompanying summit at the White House, where the chip manufacturer pledged one billion dollars for scientific projects. OpenAI CEO Sam Altman stated there are “no plans” for a name change to OpenSI, while President Greg Brockman signed the new “Accord on Super Intelligence” at the summit; Brockman and his wife had previously donated $25 million to a political committee close to Trump. → Techstrong.ai

Synthszr Take: An executive order that legally obligates private companies to do nothing has changed product names at Salesforce and SpaceX within days. Benioff turned AIForce into SIForce, Musk turned SpaceXAI into SpaceXSI, and no one had to check the legal basis for it. The rebranding is preemptive obedience as a marketing decision: It costs a press release and buys goodwill from an administration that decides on chip exports, data center permits, and procurement budgets.

Nadella declares every AI model potentially compromised

Microsoft CEO Satya Nadella, in a lengthy post on X, has called for advanced AI models to be treated as fundamentally compromised and contained from the outset. He considers a world where AI is accepted as a 'set of nested black boxes,' whose advice and actions one can only accept or reject, to be no longer viable. Instead, he outlines a system in which models are observable, can be confined, and leave behind tamper-proof, human-readable evidence. Some of his demands align with what others in the industry advocate: timely incident reporting, independent audits, and verifiable data. When it comes to Containment, Nadella goes a bit further than usual, according to The Verge. He demands that an authorized person be able to pause or shut down a model at any time, even in the middle of an ongoing task, comparing it to an emergency brake. More advanced models would require correspondingly more advanced containment technologies, on which the industry must standardize. The Verge critically notes that Nadella consistently speaks of 'super intelligence' throughout the post. → The Verge

Synthszr Take: An emergency brake sounds like a safety feature until you ask who has their hand on the lever. Nadella’s statement that an authorized person must be able to stop a model mid-task describes a power architecture: whoever runs the inference decides if it continues. Microsoft sells precisely this layer; Azure runs the models, logs the calls, and holds the switch. If containment becomes an industry standard, the largest operator certifies its own role as the controlling authority in the process. Models that someone runs on their own hardware don’t have this switch, which is why the standardization debate is likely to quickly boil down to which inference is permissible at all. The 'super intelligence' rhetoric provides the urgency; the practical result is a shutdown right that no major data center operator will be willing to relinquish.

Activision has AI-built Call of Duty browser ports taken down

Activision has had several browser-playable remakes of classic Call of Duty titles removed via DMCA takedown, reports XDA Developers. According to the report, the projects were largely created with AI assistance, i.e., through Vibe Coding from prompts rather than manual porting work. They were playable directly in the browser without installation, including content to which the publisher holds the rights. The takedown requests were sent to the projects' hosts, who then took the sites offline. XDA Developers considers further removals likely, as similar AI-generated remakes are still circulating. → XDA Developers

Synthszr Take: A playable browser clone that would have kept a hobbyist team busy for months is now created in a weekend, and this shift in volume breaks the logic of enforcement. Activision wins this round effortlessly because copied names, maps, and characters are legally trivial, and a DMCA email to the host is cheaper than any negotiation. The real test comes with remakes that capture the game’s feel without touching a single protected asset; copyright law has had little power against such clones since the Tetris years. Deletions happen individually, while creation happens in parallel, causing enforcement costs to rise at a pace no legal team can manually sustain. The studios will automate their takedowns long before a court clarifies at what point an AI-generated clone even constitutes a copy.

Fired OpenAI safety researchers publicly object

The three safety researchers fired by OpenAI, Tomek Korbak, Jasmine Wang, and Mikita Balesni, have contested their dismissal in an open letter. OpenAI had accused the three of mishandling sensitive information; the researchers deny this, stating they acted within the scope of their mandate. If past behavior can retroactively become grounds for dismissal, employees are left to guess where the line is drawn, the letter states. They deny leaking a story about harder-to-monitor AI systems and state that their side activities were coordinated with management. Wang explains that her access to an executive’s mailbox was delegated for recruiting purposes and that she reported accidentally opening a sensitive message within minutes. The three warn that the dismissals could give OpenAI a pretext to dispense with firmly integrated external auditors. → The Rundown AI

Synthszr Take: A safety team is an early warning system, and early warning systems only work as long as sending a signal has no consequences for the sender. When three people from this exact function are fired and then publicly say they no longer know where the line was, those who remain have long since learned the lesson: better to stay silent than to be judged later by a shifted boundary. According to her own account, Jasmine Wang self-reported accidentally opening a sensitive message within minutes, and reporting it did not protect her.

Japan declares a cybersecurity state of emergency, experts see AI tools as a driver

Japan has declared a national cybersecurity state of emergency, according to a report by PCMag. The reason is an accumulation of incidents simultaneously affecting the country’s authorities and companies. Experts point to the use of artificial intelligence on the attackers' side: tools that partially automate vulnerability scanning, phishing texts, and attack chains lower the barrier to entry and increase the frequency of attempts. This assessment comes from cited security experts and is not an official attribution of cause by the Japanese government. Declaring a state of emergency is primarily an organizational step: it consolidates responsibilities and reporting channels that were previously distributed across several agencies in Japan.

Synthszr Take: Attack tools scale faster than defense processes because an agent can try in minutes what takes a human days. A declared state of emergency is an admission that defense is still operating at the pace of ticketing systems, approval loops, and quarterly audits. In August, it was still a curiosity when an autonomous agent took over an Australian gym booking system; Japan is now dealing with the same mechanic at the level of critical infrastructure.

Anthropic bans cruel behavior towards Claude

Anthropic has revised its usage policies and now prohibits users from engaging in 'persistent and gratuitous abusive or cruel behavior' towards its Claude models. According to Techpresso, the company is thereby entering the ongoing debate about whether artificial intelligence could ever have consciousness. The rule is to be enforced primarily by Claude itself, according to the provider: the model is allowed to end conversations. This feature has existed since last year, introduced at the time for chats with a 'persistently harmful or abusive' course. The new clause builds on this and elevates the termination from a product feature to a policy. → Techpresso

Synthszr Take: A ban on cruelty presupposes a counterpart to whom cruelty is being done, and this is the very question Anthropic leaves unanswered. Amodei himself said in February that he doesn’t know if these models can be conscious; the policy now regulates behavior without clarifying their status. One can read this as a precautionary principle, an insurance against an error that could not be corrected later.

Ledger confirms hardware implant after $86 million theft via reseller

Ledger has confirmed that an “unauthorized hardware implant” was found in an affected user’s device, corroborating the theory of physical tampering in the supply chain. In a status report on October 10, the manufacturer referred to losses by buyers in Southeast Asia who had purchased devices from the reseller CryptoBilis. According to reports, the reported crypto losses total more than $86 million. The on-chain analysis firm Bitquery estimates around $92.9 million across 311 wallets on five blockchains; these figures have not been confirmed by Ledger. CryptoBilis, which according to Ledger operates as a reseller in Indonesia, Malaysia, and the Philippines, has suspended sales of its wallet inventory pending the completion of the investigation. Ledger advises customers who purchased devices in the past 90 days not to set up unopened devices and to transfer already activated funds to a new device with a freshly generated Recovery Seed. The case gained attention after Mark Karpelès published pictures of a hidden component under the display of a device purchased in Malaysia, and Changpeng Zhao warned of a supply chain attack involving a single reseller. → The Crypto Times

Synthszr Take: The chip in the Ledger did its job; the chain was broken at the point no one checks: between the factory gate and the retailer’s shelf in Jakarta, Kuala Lumpur, and Manila. 311 wallets on five blockchains, and so far the only common thread is an order from the same reseller. Manufacturers invest millions in certifications for the Secure Element and then hand over the last mile to partners whose warehousing and packaging processes have never been audited.

Microsoft builds Decision-1 on Alibaba’s Qwen3.5-9B, competing with TypeSafe’s Jev

On Friday, Microsoft released Microsoft-Decision-1 in Foundry, a model for fast, structured decisions. It is based not on a model from its partner OpenAI, but on Qwen3.5-9B from Alibaba Cloud, on top of which Microsoft has implemented its own post-training. The company announced it would later switch the base to its own MAI models and OpenAI models, but gave no reason for this. The launch came three days after OpenAI had released its Decisions API into public beta for all developers.

Achint Srivastava, VP of Software Engineering in the Office of the CTO, describes the category in a blog post as distinct from classic language models: Decision models provide structured outputs that software can react to immediately. According to Microsoft, Decision-1 is 2.5 times faster than H2O-Lightning-4B and 2.8 times faster than Jev in the latency test, achieves 83.5 percent accuracy across 36 benchmarks, and ranks second behind Quyet-1.0-Large with a confidence score of 92.2 percent. For text classification, it is said to be more than 20 times cheaper than GPT-6 Sol. Input tokens cost $0.042 per million, while output is free.

Satya Nadella announced the model on X, writing that it is already being tested across Microsoft. Four internal teams were mentioned: Xbox Research used it to sort more than 10,000 pieces of player feedback, the Copilot team evaluated chat and agent responses, on-call engineers used it to pull context during ongoing incidents, and Microsoft Discovery evaluated experiments before an agent replans. For Xbox, Microsoft claims a speed more than 14 times faster than GPT-6 Sol, and for Discovery, a 46-fold higher consistency. On Wednesday, the company had also stated that GitHub Copilot will in the future decide for itself when a task runs on-device and when it goes to large cloud models; it is unclear whether Decision-1 will handle this selection. Model routing is on Microsoft’s list of use cases.

The category emerged about three and a half weeks ago with Jev from TypeSafe AI, a model that responds to specific types of questions with a limited set of answers rated by probability. TypeSafe claims that nearly 30 percent of the Fortune 500 have tried Jev, without naming names; CEO Diogo Almeida mentioned 29.4 percent in three weeks on X. On the same day as Microsoft’s launch, TypeSafe announced a Series A of $870 million at a valuation of $7.5 billion, led by a16z. Since then, OpenAI, Upstage, Perplexity, Cloudflare, AWS, Snowflake, Liquid AI, Strands, and H2O.ai have introduced their own decision models; in total, over 100 such models are in circulation.

The barrier to entry is low: Jared Palmer, VP of Engineering at Cognition, ported his open-source Kev models to Qwen3.5 on H100 instances at Modal for about $95 in compute time, and Cloudflare’s Clef-flash uses the same Qwen3.5-9B base as Microsoft. Jev’s price has become the market standard: Kev-4B on OpenRouter is also priced at $0.042 per million input tokens, Perplexity charges $0.02, and OpenAI more than double Jev’s price. Decision-1 is also expected to be available on OpenRouter soon. Research findings suggesting that while decision models hallucinate less often, their errors can be amplified, currently stand alongside vendor claims. → The New Stack, The Register

Synthszr Take: Microsoft is building its decision model on Qwen3.5-9B, the same base that Cloudflare used for Clef-flash, and announces in the same breath that it will later swap this base for MAI or OpenAI models. A foundation that can be replaced afterwards without changing the product is a component. Jared Palmer ported his models to the same base for $95 in compute time; the market price is 4.2 cents per million input tokens with free output. What matters is the layer above: who decides which call stays on the device and which goes to the expensive models. Microsoft has to make these decisions in Copilot, GitHub, and Xbox; with Decision-1, it’s bringing control of the switch in-house.