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 RadarHabr AI→ original

Roskomnadzor Spent 2.27 Billion Rubles on an AI Traffic Filter, but Its Accuracy Is Undisclosed

According to the passport of its 2026 program, Roskomnadzor spent 2.27 billion rubles on a machine learning system to detect VPNs and banned websites. Previously, the filter checked traffic against lists of addresses. Now, experts suspect the model evaluates packet sizes and session durations without decrypting content. Accuracy has not been published, so teams with corporate VPNs should check their channels in advance.

AI-processed from Habr AI; edited by Hamidun News
Roskomnadzor Spent 2.27 Billion Rubles on an AI Traffic Filter, but Its Accuracy Is Undisclosed
Source: Habr AI. Collage: Hamidun News.
Listen to article
Article Fact Sheet & Key Figures
ParameterDetails
Primary sourceHabr AI ↗
Editor-in-Chief
Key metric #1TSPU equipment is installed at operators; ASBI manages it centrally
Key metric #2The contractor is not named, and the link to Astra AI from May rumors is not confirmed
Key metric #3As of November 14, 2026, there is no official statement from Roskomnadzor

On November 13, 2026, Forbes, followed by Meduza, 3DNews, and SecurityLab, reported that Roskomnadzor had deployed a machine learning-based traffic filtering system. Put simply, the agency now has a program that is supposed to tell an ordinary connection apart from one used to bypass blocks, such as a VPN. According to accounts of the document, it cost 2.27 billion rubles. All the outlets rely on the same source: the passport of the agency's digital transformation program for 2026 and the 2027–2028 planning period. If you build services, maintain corporate VPNs, or depend on cloud channels, it is worth separating what the document says from what experts have inferred.

What the passport explicitly states

According to accounts of the passport, the system runs a Russian model on Roskomnadzor's own infrastructure. It is built into the automated security system of the Russian internet segment (ASBI). ASBI centrally manages TSPU equipment, the technical countermeasure devices installed on operators' networks. Using DPI (deep packet inspection), TSPU restricts websites, slows connections, and blocks protocols.

Separate line items in the 2026 program allocate another 3.99 billion rubles for installing and upgrading TSPU and 6.2 billion rubles for equipment. The contractor and technical specifications have not been disclosed. In May, reports said the Astra AI platform from Group Astra might be used for this, but the new passport does not name the product, so the connection is unconfirmed. According to SecurityLab, the system does not use large language models.

How it may work: still a hypothesis

3DNews writes that the system is meant to find circumvention tools, including VPNs, and sites with prohibited content. SecurityLab describes a possible principle. The previous filter worked like a list check: it matched connections against known addresses, domains, and signatures. A model, however, could assess a connection based on indirect features: packet size and timing, session length, the ratio of incoming to outgoing traffic, and server characteristics. The content itself is not decrypted.

An everyday analogy: a doorman recognizes a courier not by the delivery slip but by how they walk and when they arrive. If the system works this way, changing the address or the protocol wrapper alone may not help as long as the traffic statistics remain distinctive. But this is the experts' conclusion, not a description of the implementation in the passport.

The same experts believe this is more of a predictive model than a language model: it estimates the probability that a VPN is in use from external features rather than "understanding" text.

What we don't know

  • Accuracy. Test results, the share of correct detections, and the share of false positives have not been published anywhere.
  • Impact on speed. Experts warn that at current speeds the model could become a bottleneck and add latency.
  • Place in the chain. The model may analyze traffic in real time, or, as SecurityLab suggests, it may generate signatures that TSPU then applies. According to the outlet, the second option is cheaper, but which one was chosen is unknown.
  • Impact on VPNs. Neither Roskomnadzor nor Forbes has described how the system affects specific VPN protocols.

Experts call false positives likely: corporate VPNs and cloud connections could be caught by the filter. This is a risk assessment, not documented cases.

Context and what to do

In October 2026, the Ministry of Digital Development proposed using AI to find and block VPNs, and experts already warned at that time of erroneous blocks in the early stages. The intention to build such a system first surfaced in January 2026. The program passport confirms a long-standing course rather than revealing anything unexpected.

As of November 14, 2026, there is no official statement from Roskomnadzor about this system; only the content of the passport, as relayed by Forbes, is known. The practical takeaway is straightforward. There is no reason to panic, and no reason to assume that bypassing blocks is already impossible. Product and infrastructure teams that depend on VPNs or cloud channels should measure the stability of their connections in advance and keep backup routes ready.

Two things are worth watching. First, whether independent measurements appear showing how the system affects specific protocols and latency. Second, whether the agency itself says anything about it. For now, all conclusions about the system's capabilities remain hypotheses, and the one confirmed fact is that money has been allocated to it and it is integrated into the existing TSPU infrastructure.

Story

Key Questions Answered

What the passport explicitly states?

According to accounts of the passport, the system runs a Russian model on Roskomnadzor's own infrastructure. It is built into the automated security system of the Russian internet segment (ASBI). ASBI centrally manages TSPU equipment, the technical countermeasure devices installed on operators' networks. Using DPI (deep packet inspection), TSPU restricts websites, slows connections, and blocks protocols.

How it may work: still a hypothesis?

3DNews writes that the system is meant to find circumvention tools, including VPNs, and sites with prohibited content. SecurityLab describes a possible principle. The previous filter worked like a list check: it matched connections against known addresses, domains, and signatures. A model, however, could assess a connection based on indirect features: packet size and timing, session length, the ratio of incoming to outgoing traffic, and server characteristics. The content itself is not decrypted.

What we don't know?

Accuracy. Test results, the share of correct detections, and the share of false positives have not been published anywhere. Impact on speed. Experts warn that at current speeds the model could become a bottleneck and add latency. Place in the chain. The model may analyze traffic in real time, or, as SecurityLab suggests, it may generate signatures that TSPU then applies. According to the outlet, the second option is cheaper, but which one was chosen is unknown. Impact on VPNs.

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…