Roskomnadzor allocates 2.27 billion rubles for ML-based traffic filtering through 2028
Roskomnadzor has budgeted 2.27 billion rubles for a system that uses machine learning to search traffic for banned content and ways around blocks. According to Forbes, the model is "domestic," runs on the agency's own infrastructure, and has no publicly disclosed name. Experts expect more precise blocking, but they worry about false positives and VPN failures. Teams that rely on VPN channels should plan backup routes in advance.
AI-processed from vc.ru; edited by Hamidun News
Workarounds for blocks change faster than registries of banned addresses get updated. Now dedicated funds have been allocated to this problem. According to Forbes, which cites the passport of the digital transformation program for 2026–2028, Roskomnadzor has received 2.27 billion rubles to deploy a machine learning mechanism for filtering internet traffic.
What is known about the system
The facts are few, and there is no point in speculating beyond them. The program passport indicates that the regulator uses a "domestic" mathematical model. Its name is not disclosed. The model is deployed on the agency's own infrastructure, meaning it runs on its servers rather than in a third-party cloud. The model's task is to automatically extract the needed information from large volumes of data.
Public materials do not specify what counts as "needed information." Experts interviewed by Forbes explain the system's purpose: in their assessment, the technology will help Roskomnadzor more accurately identify banned content and ways of bypassing blocks. This is their interpretation; the passport, as relayed in reporting, does not contain such wording. Phased deployment timelines, quality metrics, and the acceptable error rate are not named in the available data.
How it works, with an everyday example
A classic filter is like a security guard with a list of names: if your name is on the list, you are not let in. A machine learning model works differently. It learns from examples and estimates how closely a data flow resembles what should be filtered out. Instead of asking "is this address on the list?", it asks a different question: "does this traffic, based on its features, resemble VPN protocol operation?"
This approach has an obvious advantage. It can catch new bypass methods that do not yet appear on any list. It also has a major drawback: any statistical model makes mistakes. It may flag as prohibited something that looks similar but has nothing to do with the ban.
The same experts warn of false positives and reduced stability for corporate services and VPNs. For businesses, the risk is concrete. A working VPN between offices, an encrypted channel to a cloud database, or a connection to a partner's API may share network characteristics with what the system is trying to filter. Imagine an online store whose warehouse connects to its accounting system through an encrypted tunnel: if that tunnel starts dropping once an hour, orders get stuck.
What developers and product managers should do
While the system is only budgeted, there is no need to panic. But you can prepare now.
First, list the business-critical processes that depend on VPNs and non-standard network channels, and check what happens when the connection is unstable. Second, if you build services for Russian users, plan backup routes and monitor availability from different networks and different carriers. This costs less than investigating an outage in the middle of a failure. Third, log unexplained disconnections with timestamps and carrier information. Such data makes it easier to talk to your provider and the regulator.
Limitations and open questions
The main limitation is a lack of information. It is unknown which specific model is used, what data it was trained on, or how disputed cases will be handled. It is also unclear whether companies will have an official way to contest an erroneous block. Without these details, the real accuracy and the scale of the risks cannot be assessed.
The sum of 2.27 billion rubles over three years shows that the agency is treating machine learning as a separate expense line and is not limiting itself to a pilot. For the market, this is a signal: the network environment in which Russian services operate may become less predictable. Teams whose businesses depend on secure communication channels should prepare in advance and test their fallback scenarios before the first failures occur.
- Nov 21, 2026Roskomnadzor to Spend 2.27 Billion Rubles on AI-Based Traffic Filtering Through 2028
- Nov 21, 2026Roskomnadzor Spends 2.27 Billion Rubles on AI That Will Hunt for VPNs in Traffic
- Nov 21, 2026Roskomnadzor Spent 2.27 Billion Rubles on an AI Traffic Filter, but Its Accuracy Is Undisclosed
- Oct 9, 2026Russia launches a network of flower vending machines, and Flowwow's revenue grows to 5.7 billion rubles
Key Questions Answered
What is known about the system?
The facts are few, and there is no point in speculating beyond them. The program passport indicates that the regulator uses a "domestic" mathematical model. Its name is not disclosed. The model is deployed on the agency's own infrastructure, meaning it runs on its servers rather than in a third-party cloud. The model's task is to automatically extract the needed information from large volumes of data.
How it works, with an everyday example?
A classic filter is like a security guard with a list of names: if your name is on the list, you are not let in. A machine learning model works differently. It learns from examples and estimates how closely a data flow resembles what should be filtered out. Instead of asking "is this address on the list?", it asks a different question: "does this traffic, based on its features, resemble VPN protocol operation?" This approach has an obvious advantage.
What developers and product managers should do?
While the system is only budgeted, there is no need to panic. But you can prepare now. First, list the business-critical processes that depend on VPNs and non-standard network channels, and check what happens when the connection is unstable. Second, if you build services for Russian users, plan backup routes and monitor availability from different networks and different carriers. This costs less than investigating an outage in the middle of a failure.
Test Frontier LLMs & Autonomous Agents in One Workspace
Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.