Russia's Digital Ministry, the EU and NIST Set New Rules for AI Security and Regulation

Security Vision has published a large review of how a new system of rules, standards and defensive practices is quickly forming around AI. The focus is not only on laws but also on very practical issues: data leaks, prompt injection, model theft and control over agentic systems.

How AI is regulated

In Russia, the baseline remains Decree No. 490 of 10 October 2019 and the updated AI development strategy through 2030. Since then, the country has introduced an experimental regime in Moscow, a draft concept for AI regulation through 2030 from the Ministry of Digital Development (Mintsifry) and the AI Alliance, and an ethics code from the Central Bank for the financial market. The overall direction is clear: encourage AI adoption while demanding transparency, data protection, risk management and human oversight.

Abroad, the framework is getting stricter. The EU's AI Act is already in force, sorting systems into unacceptable, high and ordinary risk levels: from outright bans on social scoring and a number of biometric practices to mandatory logs, human oversight and cybersecurity requirements. In 2025 the US released its AI Action Plan, which emphasizes faster innovation, infrastructure and secure by design. China is tightening labeling of AI content, South Korea and Kazakhstan have passed their own laws, and the UN, UNESCO and BRICS are building an international venue for shared rules.

Standards and oversight

Alongside the laws, a layer of standards and frameworks is growing quickly. ISO and IEC have already published documents on terminology, risk management, the AI lifecycle, robustness assessment and data security. In Russia, new GOST standards correspond to them, and in the US NIST is developing the AI Risk Management Framework, which has become one of the most visible practical schemes for companies. Its logic is simple: don't debate AI in a vacuum, but build its risks into ordinary product, security and development management.

• Govern — assign owners, policies and processes for AI risks
• Map — understand context, use cases and potential harm
• Measure — test reliability, validity and the effectiveness of safeguards
• Manage — prioritize risks, reduce damage and document decisions

A separate layer of specialized frameworks is also forming: MITRE ATLAS catalogs tactics for attacking AI, OWASP compiles critical risks for LLMs and agentic systems, Google is developing SAIF, and major vendors publish their own safety and preparedness frameworks. This matters for one reason: AI security is no longer limited to protecting a server or an API. Teams now have to assess model behavior, data quality, resistance to manipulation and the consequences for the user.

Where AI gets broken

AI has its own attack surface. Large language models struggle to tell data from instructions, so malicious text on a web page, in an email or in a document can turn into a command for an agent. This drives the rise of direct and indirect prompt injection attacks, jailbreaks, extraction and data poisoning. If a system is connected to email, a CRM, file storage or external services, the cost of a mistake rises sharply: an attacker can do more than get a strange answer, and may reach corporate data or make an agent perform an unwanted action.

The second major risk zone is the data itself. Companies increasingly face Shadow AI, where employees upload internal documents to public chatbots without approval. IBM cited an estimate that 13% of surveyed companies had already reported leaks through AI. Add the theft of API keys for LLM jacking, possible model copying through distillation and the problem of hallucinations on synthetic data, and it becomes clear why grounding, zero data retention, MLSecOps and DLP are no longer optional. Another trend: attackers are themselves using agentic tools to automate hacking and lower the barrier to entry into cybercrime.

What this means

The AI market is entering a phase in which the winner is not whoever deploys a model fastest, but whoever can prove its reliability, explainability and security. For business this means one thing: AI can no longer be launched as "just another SaaS service." It needs separate access rules, data controls, audits of integrations and regular testing against attacks.