Free Intensive Nov 21–22Free 2-day hands-on Vibecoding Intensive on Nov 21–22 at 16:00 MSK (Sat–Sun) · Live Cohort 3 starts Nov 23Join Free →Cohort 3 (Nov 23)
ReleaseInsight · On Radar3DNews AI→ original

Roskomnadzor Spends 2.27 Billion Rubles on AI That Will Hunt for VPNs in Traffic

Roskomnadzor has spent 2.27 billion rubles on an AI model that will search traffic for VPNs and banned content. It is built into ASBI, the Runet security system, alongside TSPU, the operator-side equipment that filters traffic. Experts believe the model does not read content. Instead, it estimates the probability of VPN use from external traffic characteristics. Companies that rely on corporate VPNs should therefore keep a backup communication channel ready in advance.

AI-processed from 3DNews AI; edited by Hamidun News
Roskomnadzor Spends 2.27 Billion Rubles on AI That Will Hunt for VPNs in Traffic
Source: 3DNews AI. Collage: Hamidun News.
Listen to article
Article Fact Sheet & Key Figures
ParameterDetails
Primary source3DNews AI ↗
Editor-in-Chief
Key metric #1In 2026, operators received 3.99 billion and 6.2 billion rubles for TSPU equipment.
Key metric #2In May, reports said the agency planned to use Astra AI, a neural network from Group Astra.
Key metric #3Risks include false positives and traffic delays caused by a "bottleneck."

Roskomnadzor has spent 2.27 billion rubles on a machine learning model designed to find VPNs and banned content in internet traffic. The model has been built into ASBI, the automated security system for the Russian internet segment, where it complements TSPU. TSPU stands for technical countermeasures against threats: equipment installed at telecom operators that filters the traffic passing through it. If your product depends on a stable connection, such as a corporate VPN, remote access to infrastructure, or SaaS for customers in Russia, this news affects you directly. The network environment has gained another unpredictable variable.

What Was Actually Built

The AI traffic filtering is mentioned in the passport of the digital information program for 2026, with planning extending into 2027 and 2028. According to the source, the system is based on a domestic AI model that the agency has deployed within its own infrastructure.

Plans to create such tools were reported in January. In May, it became known that the agency intends to use Astra AI, a neural network developed by Group Astra, for this purpose. Separately, this year operators have been allocated 3.99 billion rubles from primary funding and 6.2 billion rubles from supplementary funding to install new TSPU units and upgrade old ones. Together, that comes to about 10.2 billion rubles in equipment spending.

How It May Work

The material contains no official technical details, so what follows are experts' assumptions. In their view, this is not a large language model like ChatGPT but a predictive model. It does not read packet contents. Instead, it looks at external traffic characteristics: how the traffic is structured and how the connection behaves. Based on these signals, it estimates the probability that a user is running a VPN.

A everyday analogy: a security guard does not look inside your bag, but notices that you behave like people who were detained before. If the filter works this way, common tactics such as traffic obfuscation and address switching will help less. VPN services may degrade in quality even when they disguise their traffic.

AI could also play a second role, helping with signatures. DPI systems identify and filter traffic without decrypting its contents, but they depend heavily on up-to-date signatures, the patterns used to recognize traffic. When those patterns go stale, filtering effectiveness drops. Owners of blocked websites often create mirrors, and banned material resurfaces on other resources. Checking such a volume by hand is impossible. Machine learning may be used to match such material against samples rather than relying only on domain names. These are hypotheses, not confirmed functions.

Risks and What Practitioners Should Do

The source explicitly names limitations. First, the AI will not deliver perfect accuracy: errors and false positives will remain. Second, legitimate corporate services may end up blocked, not only consumer "entertainment" VPNs but also companies' work tunnels. Third, at modern connection speeds, the AI filter could become a "bottleneck," a source of delays and user frustration.

What teams building products and infrastructure should do:

  • Build connection degradation into your availability plans. If a service relies on a VPN or non-standard protocols, you need a fallback channel and client-side monitoring.
  • Do not assume traffic obfuscation solves the problem. If the model evaluates behavioral signals rather than content, encryption alone does not guarantee that a connection will not be treated as a circumvention attempt.
  • Log incidents. If a corporate service suddenly fails for some operators, logs will help you distinguish a false positive from your own error.
  • Track news about TSPU and signatures. They change, and network behavior changes with them.

The scale of the spending shows that this direction is a priority: AI filtering is written into a three-year program. How accurate it will be remains unknown, because no public quality metrics exist. The practical task is to watch real-world service stability and build in margin now.

Story

Key Questions Answered

What Was Actually Built?

The AI traffic filtering is mentioned in the passport of the digital information program for 2026, with planning extending into 2027 and 2028. According to the source, the system is based on a domestic AI model that the agency has deployed within its own infrastructure. Plans to create such tools were reported in January. In May, it became known that the agency intends to use Astra AI, a neural network developed by Group Astra, for this purpose.

How It May Work?

The material contains no official technical details, so what follows are experts' assumptions. In their view, this is not a large language model like ChatGPT but a predictive model. It does not read packet contents. Instead, it looks at external traffic characteristics: how the traffic is structured and how the connection behaves. Based on these signals, it estimates the probability that a user is running a VPN.

What are the key details on Risks and What Practitioners Should Do?

The source explicitly names limitations. First, the AI will not deliver perfect accuracy: errors and false positives will remain. Second, legitimate corporate services may end up blocked, not only consumer "entertainment" VPNs but also companies' work tunnels. Third, at modern connection speeds, the AI filter could become a "bottleneck," a source of delays and user frustration. What teams building products and infrastructure should do: Build connection degradation into your availability plans.

ZK
Zhemal Khamidun — CPO at Alpina Digital, product manager of AlpinaGPT. Building AI products since 2019. Hamidun News is a pet project that grew into a production aggregator with 6,000+ readers.
✦ Unified AI Workspace · Nomad

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.

Try Nomad AI →Launch Nomad Agent
What do you think?
Loading comments…