Standards and Expertise
What standards in the field of artificial intelligence are, which of them are already in force, what the Institute applies in its work, and what it is still only considering.
What is in this section
Five short answers: why standards are needed at all, what kinds exist, what is already in force in Russia and abroad, what we apply in our work, and what ISO/IEC 42001 really gives you. Each item jumps straight to its block — no need to scroll the whole page.
- Why standards are needed — what a standard is and how it differs from a certificate and from a methodology.
- What kinds of standards there are — five levels, from international ISO to an in-house regulation.
- Standards map — what is already in force: Russian GOST R and PNST on AI, international and regional documents, with status and dates.
- Our position and practice — what we apply at work, what we are still only considering, and why we are in no hurry.
- ISO/IEC 42001 in plain words — what a certificate confirms, what it does not give, and how it connects to our testing.
Why standards are needed
A standard is an agreement on how to do things and how to verify them. The law answers the question of what must not be done; a standard answers the question of what is accepted practice and how to prove that something has been done correctly. In engineering such a division is familiar: the strength of a bridge is not assessed by eye, there is a method of calculation and a method of testing. In the field of artificial intelligence this habit does not yet exist — and this is one of the reasons why the conversation about AI safety so often remains a conversation about intentions.
A standard is often confused with a certificate. A certificate is a document that someone has issued following an assessment; a standard is what the assessment is carried out against. A methodology guide is one organisation's own document; a standard is backed by a coordination procedure and by recognition on the part of the parties. A recommendation, for example a risk management framework, obliges no one to anything, but it is convenient to refer to it while there is no standard in that field.
The usefulness of a standard is common to all, while each party's gain is its own. For the customer it provides the language of a contract: the requirement "complies with the standard" can be verified, unlike the wish "make it safe". For the developer it marks the boundary of what is sufficient: working to a standard is cheaper than proving every time that a solution is reasonable. For the regulator it is a support that does not require a new law: a standard changes faster than a law and outlives changes in technology.
In artificial intelligence everything is more difficult. The field changes faster than standards are written: while a document is going through discussion, the next generation of models appears. There are no units of measurement either: the strength of concrete has a number, whereas the robustness of a model to manipulation has so far only a set of scenarios and an agreement on what is to be considered a failure. That is why a standard in such a field is not the truth, but a way of recording what exactly has been verified and what has remained outside the scope of verification.
What kinds of standards there are
Standards differ not only in content but also in level. The level determines who recognises the document and what it gives in practice.
| Level | Who issues it | Example | What it is useful for |
|---|---|---|---|
| International | ISO, IEC, IEEE, ITU | ISO/IEC 42001 — Artificial intelligence. Management system | recognised in different countries, convenient for working with a foreign partner |
| Regional | CEN-CENELEC, the European Union | EN 18286:2026 — the first harmonised standard for the EU Regulation on artificial intelligence | turns the requirements of the law into verifiable procedures |
| National | Rosstandart and technical committees | GOST R and PNST on artificial intelligence | the basis for procurement, supervision and expertise within the country |
| Industry and departmental | the regulator of the industry, associations of organisations | requirements for medical systems with AI | takes into account the risk and the practice of a particular industry |
| Corporate | the organisation itself | an internal regulation on the use of models | is written for one's own processes and is usually stricter than a national one |
Standing apart are documents that are not standards yet are used in AI work perhaps even more often: the NIST AI RMF risk management framework, the MITRE ATLAS catalogue of attack tactics and techniques, the OWASP list of typical vulnerabilities of applications built on large language models. Legally these are recommendations and catalogues: they oblige no one. But it is these that reports refer to, because standards have not yet covered this field.
How this is arranged in Russia
National standards are issued by Rosstandart, and they are developed by technical committees — standing platforms where developers, customers, researchers and regulators meet at one table. For artificial intelligence the relevant committee is TK 164 "Artificial Intelligence": it was established by an order of Rosstandart in 2019, works in five subcommittees (healthcare, data, the road transport sector, railway transport, industry) and performs the functions of Russia's national body in the relevant ISO and IEC committee.
The path of a document looks approximately as follows:
- Application. An organisation proposes a topic and justifies why a standard is needed. The application is considered by the technical committee.
- Working group. The committee brings together developers and writes a draft: what exactly is standardised, which terms are used, how conformity is verified.
- Public discussion. The draft is put out for comments; objections are examined on their merits, and not according to the departmental affiliation of their author.
- Approval and entry into force. The standard receives a designation and a date of entry into force; from that moment it can be cited in a contract.
Besides GOST R there is also a preliminary national standard — the PNST. It is issued when a field changes quickly and it is pointless to wait for a full coordination cycle: a PNST is in force for a limited period, gives practice a reference point, and is then reworked into a GOST R or withdrawn. For artificial intelligence it is a working tool, and not an exception.
Standards map: what is already in force
Below is not an exhaustive list, but what is relevant to our work: AI management, risk management, testing, quality and data. The designations are clickable and lead to the official catalogue entry.
National standards of Russia
| Designation | Title | Subject and status |
|---|---|---|
| GOST R ISO/IEC 42001-2024 | Artificial intelligence. Management system | The Russian version of ISO/IEC 42001. Requirements for AI management in an organisation: roles, processes, risk assessment, continual improvement. Voluntary certification of an AI management system is built around it. In force from 01.01.2025. |
| GOST R 71476-2024 | Artificial intelligence. Artificial intelligence concepts and terminology | A basic vocabulary: what counts as an AI system, what training, inference and autonomy are. The Russian version of ISO/IEC 22989:2022. In force from 01.01.2025. |
| GOST R 71539-2024 | Artificial intelligence. AI system life cycle processes | The stages of work with a system — from concept and data to decommissioning (ISO/IEC 5338:2023). A framework that shows what has been lost in a project. In force from 01.01.2025. |
| GOST R 59276-2020 | Artificial intelligence systems. Methods for ensuring trust. General provisions | One of the first national standards in the series: what it means to trust a system and by what methods trust is confirmed. In force from 01.03.2021. |
| GOST R 59898-2021 | Quality evaluation of artificial intelligence systems. General provisions | How to evaluate the quality of a system and which evaluation methods are used for this; a basis for acceptance criteria. In force from 01.03.2022. |
| GOST R 72514-2026 | Artificial intelligence. AI system impact assessment | How to assess the impact of a system on people and groups — before and during operation (ISO/IEC 42005:2025). In force from 01.05.2026. |
| GOST R 72515-2026 | Artificial intelligence. Transparency taxonomy of artificial intelligence systems | What types of transparency and disclosure of information about a system exist (ISO/IEC 12792:2025). In force from 01.05.2026. |
| GOST R 71484.1-2024 — 71484.4-2024 | Artificial intelligence. Data quality for analytics and machine learning | Four parts: terminology, data quality metrics, management requirements and process structure (the ISO/IEC 5259 series). The link leads to Part 1. In force from 01.01.2025. |
| GOST R 71752-2024 | Artificial intelligence. Technical specification. Requirements for the content | What a technical specification for an AI system must contain. Useful to both parties — the customer and the developer. In force from 01.01.2025. |
| GOST R ISO/IEC 24029-2-2024 | Artificial intelligence. Assessment of the robustness of neural networks. Part 2. Methodology for the use of formal methods | Robustness of models to input distortions; Part 1 is GOST R 70462.1-2022. Closest to the topic of adversarial robustness. In force from 01.01.2025. |
| PNST 839-2023 | Artificial intelligence. Bias in artificial intelligence systems and artificial intelligence aided decision making | A preliminary standard on bias in models (ISO/IEC TR 24027:2021). In force from 01.01.2024. |
| PNST 1046-2026 | Artificial intelligence in critical information infrastructure. General provisions | The use of AI at significant facilities of critical information infrastructure. In force from 01.04.2026. |
The table does not show everything: the official Rosstandart catalogue lists about 150 documents in force under the “Artificial intelligence” heading — there are series on testing algorithms in transport, on medical systems, on education, on synthetic data. Here we have collected what relates to management, risk, quality and data.
An important detail that a customer would do well to know: some of the preliminary standards have already been withdrawn. The terminology standard PNST 553-2021 has been replaced by GOST R 71476-2024, and the preliminary risk management standard translated from ISO/IEC 23894 (PNST 776-2022) has been withdrawn, and there is no replacing national standard yet. While there is none, in our work we have to rely on the international original and explain this in the report.
International and foreign documents
| Document | What it is | Where adopted and what follows from it |
|---|---|---|
| ISO/IEC 42001:2023 | Artificial intelligence management system | Adopted in Russia as GOST R ISO/IEC 42001-2024. The main document on which the management part of the audit relies. |
| ISO/IEC 23894:2023 | Management of risks related to artificial intelligence | It was translated as PNST 776-2022; that preliminary standard has been withdrawn. There is no national replacement. |
| ISO/IEC 22989:2022 | Artificial intelligence concepts and terminology | Adopted as GOST R 71476-2024. |
| ISO/IEC 5338:2023 | AI system life cycle processes | Adopted as GOST R 71539-2024. |
| ISO/IEC 42005:2025 | AI system impact assessment | Adopted as GOST R 72514-2026. |
| ISO/IEC TR 24027:2021 | Bias in artificial intelligence systems | Adopted as the preliminary standard PNST 839-2023. |
| ISO/IEC 38507:2022 | Governance implications of the use of artificial intelligence by organisations | Adopted as the preliminary standard PNST 843-2023. |
| ISO/IEC 42006:2025 | Requirements for bodies providing audit and certification of an artificial intelligence management system | A separate document is needed because the management system standard does not say who is entitled to verify it. We have not come across bodies accredited under this standard in Russia: an AI management system certificate is today obtained under an international scheme. |
| NIST AI RMF 1.0 | The Artificial Intelligence Risk Management Framework, 2023; in 2024 a profile for generative AI was issued for it | A voluntary framework, not a standard: it says outright that use is optional. It has no regulatory force in Russia, but it is referred to in reports. |
| NIST AI 100-2 | A taxonomy and terminology of adversarial attacks on machine learning systems | Not a standard: a common language for describing attacks — data poisoning, model evasion, privacy attacks, misuse. A convenient basis for a testing report. |
| MITRE ATLAS and OWASP Top 10 for LLM and the OWASP Top 10 for Agentic Applications | A catalogue of attack techniques against AI systems and industry lists of risks for applications built on large language models and for agentic systems | They are not standards and have no legal force; they are updated faster than standards and are therefore useful as working material when planning tests. The list for agentic systems is especially close to us: multi-agent loops are our subject. |
Some international standards have not been adopted in Russia. They may be applied voluntarily, and to refer to them in a contract is a matter for the parties: such a reference becomes a term of the transaction, but does not replace a national standard where conformity is mandatory by law.
Our public position on the themes
Below is our public position on these themes. It is based on open sources; we are ready to discuss it publicly and to refine it as practice develops.
| Theme | Our position in one line |
|---|---|
| Regulation of general-purpose models | Obligations should attach to measurable behaviour and to the ability to investigate it, and not only to the size of the model. |
| Labelling of generated content | It is the artefact that must be labelled, and its provenance that must be preserved; a label without provenance cannot be verified. |
| Disclosure of incidents | An incident that cannot be counted cannot be regulated either; a common recording format is a precondition for any serious oversight. |
| Agent autonomy | The authority of an agent should be bounded, logged and revocable, and the log of actions should be an object of audit. |
| Evaluation of safety claims | A claim of safety is meaningful only together with the set of scenarios against which it was tested. |
| Publication of findings | Findings are in the public interest; attack payloads only under responsible disclosure. |
What we have today
The organisation is not registered yet, and we do not want to look more solid than we are. Here is what actually exists.
Behavioural audit protocol
An internal document: the scope of verification, the roles of agents, metrics, ethical constraints. Draft 0.1 — a working draft, not published.
Test bench
Our own environment on our server: scenarios, runs, a log of results. For now this is a tool for our own research, not a service.
Public position
Six themes on which we have a formulated opinion — the table above. This too is part of the work: a position that can be discussed and contested.
Analysis: FSTEC Order No. 117 vs the AI Act
Verbatim quotes from the order and the AI Act, five lines of comparison and a practical conclusion for state customers and providers. Read the article.
What is still only under consideration
These lines of work are stated as plans. The status of each changes only after a real step, and we write about it plainly.
National standardisation
We are preparing an application to join the technical committee for artificial intelligence (TK 164), where the national standards in the AI series are developed. The application has not been filed yet; we will publish the status when it changes.
Expert review of draft acts
We plan to send written expert opinions on drafts of normative legal acts through the official public discussion procedure and to publish them together with the submission.
Methodological documents
Internal and public methodology: how to limit the scope of a behavioural audit, how to set acceptance criteria for an AI system, how to record an incident so that it can later be counted.
Corporate standards
For customers who need an internal regulation rather than a research paper: the document is written around the customer's processes, with control points and a designated owner.
How we apply standards in our work
- Management framework. We take the standard for an AI management system as the skeleton of the report: who is responsible for the system, what is recorded, how it is verified, what happens after an incident.
- Risk management. Risks are listed before testing and tied to specific scenarios — otherwise the report turns into a list of platitudes.
- Testing and quality. Acceptance criteria are formulated before work begins, and are not fitted to the result that has been obtained.
- Structure of the report. An item of the document — what was verified — what was found — what was not verified and why. The last column is mandatory for us.
Primary sources and citation
- The texts of national standards and the official catalogue entries are on the Rosstandart portal (protect.gost.ru). The links in the tables above lead to the catalogue entries: they show the status, the date of entry into force, the order and the developer. The catalogue of standards in force in the field of “Artificial intelligence” is on the Rosstandart website.
- International standards are in the ISO catalogue (iso.org), and documents on AI risk management are on the NIST website (nist.gov).
- We give the designations, titles and short quotations of standards with a reference to the official source. We do not reproduce the full text of a standard: it is protected as an object of copyright, and it is purchased from the rights holder. Short quotation for information purposes is lawful (Article 1274 of the Civil Code of the Russian Federation).
Expert products
Expert opinion
A written assessment of a draft act, a technical specification or an architecture: with the reasoning shown and the limits of the assessment stated explicitly.
Threat model
Assets, actors, attack paths and protective measures for a specific AI system, tied to the customer's processes.
Methodology
A repeatable procedure: steps, roles, artefacts, acceptance criteria and a checklist that a third party can verify.
ISO/IEC 42001: what it is and what it does not give
ISO/IEC 42001:2023 is a standard for an AI management system (AIMS). In structure it is a relative of ISO 27001, only about AI: policy, roles and responsibilities, AI risk assessment, model and data lifecycle management, supplier management, incident handling, continual improvement. In plain words, it answers “how does the organisation manage AI”, not “is this particular model any good”.
In Russia it applies as GOST R ISO/IEC 42001-2024, adopted by Rosstandart order No. 1549-st of 28.10.2024 and in force since 01.01.2025. It can be cited in documents and on a website with no certification at all: it is a national standard, not a quality mark.
- It does not confirm the quality of a specific system. The management process is certified, not the model: a typical certificate states that it does not authorise labelling of products or services.
- It does not replace the AI Act. The standard is adopted as EN ISO/IEC 42001:2026 but is not harmonised: the only harmonised norm is EN 18286:2026, cited in the Official Journal of the EU on 14.10.2026, and the presumption of conformity covers only the Article 17 requirements it addresses.
- It is not an impact assessment. An impact assessment is a separate document under ISO/IEC 42005; 42001 does not replace it.
- It produces no product-level evidence. The CEDAR-42001 study found 89.9 % of audit lines conformant, but only 34.3 % reached a basic high-assurance level.
How this connects to our work
We do not certify and do not issue opinions on conformity with 42001 — that is the role of an accredited body. We supply what a management-system audit does not produce.
| What we supply | What it gives the client |
|---|---|
| System test results | proof that the system behaves as its documents say |
| Metrics for LLM01 and LLM03 | numbers instead of “we assessed the risks” |
| Protocols before and after the run | reproducibility for the auditor |
| Mapping to 42001 controls | a bridge from our findings to their management system |
We map along the OWASP GenAI Crosswalk v4.0.0 skeleton: every LLM01–LLM10 is already matched there to ISO 42001 controls — A.6.2.3 (system security), A.7.2 and A.7.3 (data and provenance), A.10.1 (suppliers), Cl.6.1 (risk assessment).
If your organisation is building an AI management system under ISO/IEC 42001, we supply the evidence that management-system audits do not produce: system test results, metrics and protocols. We do not certify or opine on conformity — that is the role of an accredited certification body.
Primary sources: ISO/IEC 42001:2023; GOST R ISO/IEC 42001-2024 (Rosstandart order No. 1549-st of 28.10.2024); ISO/IEC 42006:2025; EN 18286:2026 (Official Journal of the EU, 14.10.2026); OWASP GenAI Crosswalk v4.0.0 (28.08.2026); CEDAR-42001 study (arXiv 2606.21276).
Key points
- The Institute states its public position on six topics, from general-purpose model regulation to labelling of generated content.
- The standards map is arranged in five levels, from international (ISO/IEC 42001) to corporate.
- In its work the Institute relies on four Russian national standards (GOST R) of the AI series.
- The application to join TC 164, where national AI standards are drafted, has not been filed yet — we say so plainly.
- The split is simple: a standard answers “how to verify”, the law answers “what is forbidden”.
What to remember
- A standard covers “how to verify”; the law covers “what is forbidden”.
- GOST R is the first contour, international documents the second, for cross-checking.
- Application is proven by practice: the standards map and a public position on six topics.