Habr launches AI security series: 4 jurisdictions and 4 defense methodologies

Habr AI has published the first part of a series on the security of AI systems in 2026. The article covers the regulatory frameworks of four jurisdictions (Russia, China, the EU and the US) and four defense methodologies: STRIDE, OWASP, AI-SAFE and SAIMM. The authors call this the baseline body of knowledge that every AI security specialist needs. The announcement does not analyze specific regulations, so what follows covers what the authors state and the context in which it matters.

Machine learning security was long a niche topic. A few research groups worked on it, about a dozen papers appeared per year, and there were almost no practical tools. That changed when LLMs and agentic systems reached production. A model now calls tools, reads external data and makes decisions on a user's behalf. The attack surface has grown, and companies have ended up answerable for the consequences.

Four jurisdictions, four logics

The article compares regulation in Russia, China, the EU and the US. The approaches differ markedly, as is clear from the broader context.

• The EU built a horizontal AI law (the AI Act) with tiers by risk level and separate obligations for developers of general-purpose models. The timelines for applying some of its requirements have been discussed and adjusted, so check current dates against the primary source.
• China regulates individual classes of systems: generative AI, algorithmic recommendations and labeling of synthetic content. This is a set of targeted requirements, not a single law.
• The US relies on a mix of federal frameworks such as NIST AI RMF, sector-specific requirements and individual state laws.
• Russia is developing its own regulatory and standardization base. The announcement does not make clear which acts and standards the authors consider key.

For a company operating in several markets, this means several parallel compliance tracks. There is no single standard yet.

Methodologies: what comes from abroad, what exists at home

The second half of the article is devoted to defense methodologies. STRIDE is a classic threat model that predates LLMs and is applied to AI systems as a way to structure risks. OWASP is known for its practical vulnerability lists, including for applications built on large language models. The announcement only names the Russian AI-SAFE and SAIMM: it gives no structure, no maturity criteria and no scope. Their strengths and weaknesses can be judged only after reading the full text.

The combination itself is interesting. The Western methodologies provide a common language for discussing threats. The Russian ones, judging by the available information, are tied to local regulation and certification practice. A specialist will need to work with both layers.

What teams can do right now

While the series is not yet complete, there are several steps that don't require waiting.

1. Build a registry of the company's AI systems: which models and agents are running, with what data and access rights. 2. Determine which jurisdictions each system falls under. The criterion is not where the company is registered, but where the users and data are. 3. Run STRIDE threat modeling for the most critical agentic scenarios, especially those where the model can act in external systems. 4. Map the identified risks to the OWASP lists for LLM applications and check which controls are already in place. 5. Wait for the breakdown of AI-SAFE and SAIMM and assess how well they fit your internal processes and regulatory requirements.

By the authors' intent, the series is aimed at practitioners. The first part sets the framework, and the value of the whole series will become clear once the breakdowns of specific attacks and controls appear.