FSTEC Order No. 117 vs the AI Act: what is actually being compared

The two documents are almost always placed side by side as two answers to one question. The questions differ: the Russian order protects information in state systems, the European regulation governs a market. Below are verbatim quotes and five lines along which the difference becomes visible.

Why compare them at all

In discussions of AI regulation, “FSTEC No. 117” and “the AI Act” are almost always mentioned together: one document Russian, the other European, both about artificial intelligence. That proximity invites the question “which is stricter”, and it has no answer, because the documents regulate different things.

We took both texts verbatim and compared them along five lines: who is responsible, what happens with data, what is required of behaviour, how trust is demonstrated, and what neither document contains. The conclusion is short: they are not opponents but different regimes, and confusing them is dangerous in both directions — whether one looks for European rights protection in Order No. 117, or applies the AI Act to a Russian state system.

What exactly is being compared

FSTEC Order No. 117 of 11 April 2025 approves the Requirements for the protection of information in state information systems and in other information systems of state bodies, state unitary enterprises and state institutions. It has been in force since 1 March 2026 and replaced Order No. 17 of 2013. Artificial intelligence appears in four paragraphs: 34 “t” (AI as a separate protection measure), 49 (trusted AI technologies in monitoring), 60 (protection when AI is used) and 61 (rules for interaction with AI services).

Regulation (EU) 2024/1689, known as the AI Act, is a document of a different nature. It regulates a market: obligations are distributed between providers and deployers, and their strictness depends on the level of risk. It entered into force on 1 August 2024; most of it applies from 2 August 2026, the articles on general-purpose models from 2 August 2025, and Article 6(1) from 2 August 2027 (Article 113).

What follows is organised by line. Quotes are given in the original; where the original is Russian, our translation follows.

Who is responsible

In Order No. 117 there is one addressee: the operator (holder of information), that is, the organisation that operates a state system. There is no “model developer” role among the obligations — the developer is mentioned exactly once, as the recipient of a prohibition in paragraph 60.

The AI Act has two addressees with different duties. The provider answers for the system, the documentation and conformity; the deployer answers for use: human oversight, monitoring, incident reporting (Article 26). The same code may be a high-risk system for one and an ordinary one for another: everything depends on its intended purpose.

This is not a terminological detail. In Russia the party that uses AI is responsible; in Europe responsibility runs along the chain from developer to operator, and every link has its own set of duties.

Data: prohibition versus disclosure

This is the sharpest divergence. Paragraph 60 (second subparagraph) directly prohibits transferring restricted information to the developer of the model — and separately names what such a transfer usually aims at:

«Не допускается передача лицу, разработавшему модель искусственного интеллекта, информации ограниченного доступа, содержащейся в информационных системах, в том числе для улучшения функционирования модели искусственного интеллекта».

Translation: the transfer to the person who developed the AI model of restricted-access information contained in information systems is not permitted, including for the purpose of improving the functioning of the AI model.

FSTEC Order No. 117, paragraph 60, second subparagraph

The AI Act acts in the opposite way on data: it requires disclosure. A provider of a general-purpose AI model must have a copyright compliance policy and must publish a “sufficiently detailed summary” of the content used for training:

“Providers of general-purpose AI models shall: … (c) put in place a policy to comply with Union law on copyright and related rights …; (d) draw up and make publicly available a sufficiently detailed summary about the content used for training of the general-purpose AI model, according to a template provided by the AI Office.”

AI Act, Article 53(1), points (c) and (d)

For high-risk systems this is supplemented by data governance: the origin of data, and for personal data the original purpose of collection, plus an assessment of possible biases   (Article 10).

“Those practices shall concern in particular: … (b) data collection processes and the origin of data, and in the case of personal data, the original purpose of the data collection; … (f) examination in view of possible biases …”

AI Act, Article 10(2), points (b) and (f)

So in one case giving data away is prohibited, in the other telling where the data came from is mandatory. The logic differs: protection of information versus transparency of provenance. Neither approach is “stricter” — they protect different values.

Behaviour: templates versus risk management

Paragraph 61 sets requirements at the moment of interaction: templates for requests and answers, permitted topics, statistical criteria for detecting unreliable answers and a reaction to them. These are operational rules — something that can be checked at the moment of an   answer.

«При взаимодействии пользователей … с сервисами на основе искусственного интеллекта посредством направления запроса и получения ответа должны быть: … в) разработаны статистические критерии для выявления недостоверных ответов искусственного интеллекта …; г) обеспечено реагирование на недостоверные ответы искусственного интеллекта посредством ограничения области принимаемых решений и (или) реализации функций информационной системы на основе недостоверных ответов искусственного интеллекта».

Translation: when users interact with AI-based services by sending a request and receiving an answer, there shall be: … (c) statistical criteria for detecting unreliable answers of the AI, for subsequent collection and analysis of such answers; (d) a response to unreliable AI answers by limiting the scope of decisions taken and/or the functions of the information system performed on the basis of unreliable AI answers.

FSTEC Order No. 117, paragraph 61, points (c) and (d)

The AI Act proceeds from risk and from process: a risk management system across the whole lifecycle (Article 9), accuracy, robustness and cybersecurity (Article 15), human oversight (Article 26), transparency and marking of generated content (Article 50), and for models with systemic risk — adversarial testing and incident reporting (Article 55).

“High-risk AI systems shall be designed and developed in such a way that they achieve an appropriate level of accuracy, robustness, and cybersecurity, and that they perform consistently in those respects throughout their lifecycle.”

AI Act, Article 15(1)

“Deployers shall assign human oversight to natural persons who have the necessary competence, training and authority, as well as the necessary support.”

AI Act, Article 26(2)

The difference is in horizon: Order No. 117 checks the answer, the AI Act checks the process that produced it. In practice this means the Russian requirements are closer to what we call guardrails, and the European ones to a management system.

Trusted technologies versus conformity assessment

Order No. 117 requires that “trusted artificial intelligence technologies” be included in the system (paragraph 61) and permits their use in monitoring (paragraph 49). But what counts as trusted is not defined: the term comes from the National AI Strategy and remains open, which leaves room both for interpretation and for abuse.

The AI Act replaces the question of trust with a procedure: conformity assessment, declaration of conformity, CE marking, registration in the EU database, and — until harmonised standards appear — codes of practice (Article 53(4)). Trust here is not a property of a technology but the result of an assessment that has an executor and a document.

Comparison along five lines

QuestionFSTEC Order No. 117AI Act
Who is responsibleoperator (holder of information) of a state system provider and deployer, with different duties
Scopestate information systems and systems of state organisations the EU market, risk-based approach
Dataprohibition on transferring restricted-access information to the model developer, including for improving the model's functioning disclosure of training data, copyright policy, data governance for high-risk systems
Behaviourrequest and answer templates, permitted topics, criteria for unreliable answers, reaction risk management, accuracy and robustness, human oversight, transparency
Trust“trusted AI technologies”, undefined conformity assessment, CE marking, codes of practice

What neither document contains

  • An answer to “may AI decide for a human” — that is left to sectoral and constitutional law, and neither document substitutes for it.
  • Order No. 117 does not have transparency, marking of generated content or the protection of human rights: it is about protecting information, not rights.
  • The AI Act does not have a direct prohibition on training a model with a customer’s data — that is governed by the GDPR and by contract; nor a requirement of “domestic” technologies; nor an analogue of the Russian list of information-security measures.
  • Neither defines “trusted AI” in a way that could be applied in an assessment without further documents.

What follows from this

For a state customer in Russia, Order No. 117 is mandatory, and its AI requirements must go into the specification and acceptance — otherwise a system is formally protected while the requirements for artificial intelligence are not met. For a provider on the EU market, the AI Act applies, with risk assessment, documentation and conformity procedures.

Setting them “against” each other is a matter of polemics: legally they are two different regimes, and neither repeals the other. The practical conclusion for us is simple: with state systems we check against Order No. 117, with a foreign partner we speak in terms of the AI Act, and we write the report so that it can be read from both sides.

Sources and caveats

  • FSTEC Order No. 117 of 11 April 2025 — official publication 0001202506170011 of 17 June 2025; registered with the Ministry of Justice on 16 June 2025, No. 82619. Quotes verified against the scanned publication.
  • Regulation (EU) 2024/1689 (AI Act) — Official Journal, CELEX 32024R1689. Quotes in the original language.
  • This is an analytical review, not a legal opinion. Before using it in a contract or in acceptance, a lawyer and the current edition of the documents must be consulted.

Page updated 24 October 2026 · site build 2026.11.13 · ANO «ISAIS»